Received November 8, 2019, accepted November 29, 2019, date of publication December 4, 2019,

date of current version December 23, 2019.

Digital Object Identifier 10.1109/ACCESS.2019.2957698

Battery Energy Storage Models

for Optimal Control
DAVID M. ROSEWATER 1 , (Member, IEEE), DAVID A. COPP 2 , (Member, IEEE),
TU A. NGUYEN 1 , (Senior Member, IEEE), RAYMOND H. BYRNE 1 , (Fellow, IEEE),
AND SURYA SANTOSO 3 , (Fellow, IEEE)
1 Sandia National Laboratories, Albuquerque, NM 87185, USA
2 Department of Mechanical and Aerospace Engineering, University of California at Irvine, Irvine, CA 92697, USA
3 The Department of Electrical and Computer Engineering, University of Texas at Austin, Austin, TX 78712, USA
Corresponding author: David M. Rosewater ([email protected])
This work was funded by the U.S. Department of Energy, Office of Electricity, through the Energy Storage Program.

ABSTRACT As batteries become more prevalent in grid energy storage applications, the controllers that
decide when to charge and discharge become critical to maximizing their utilization. Controller design for
these applications is based on models that mathematically represent the physical dynamics and constraints
of batteries. Unrepresented dynamics in these models can lead to suboptimal control. Our goal is to
examine the state-of-the-art with respect to the models used in optimal control of battery energy storage
systems (BESSs). This review helps engineers navigate the range of available design choices and helps
researchers by identifying gaps in the state-of-the-art. BESS models can be classified by physical domain:
state-of-charge (SoC), temperature, and degradation. SoC models can be further classified by the units they
use to define capacity: electrical energy, electrical charge, and chemical concentration. Most energy based
SoC models are linear, with variations in ways of representing efficiency and the limits on power. The
charge based SoC models include many variations of equivalent circuits for predicting battery string voltage.
SoC models based on chemical concentrations use material properties and physical parameters in the cell
design to predict battery voltage and charge capacity. Temperature is modeled through a combination of
heat generation and heat transfer. Heat is generated through changes in entropy, overpotential losses, and
resistive heating. Heat is transferred through conduction, radiation, and convection. Variations in thermal
models are based on which generation and transfer mechanisms are represented and the number and physical
significance of finite elements in the model. Modeling battery degradation can be done empirically or based
on underlying physical mechanisms. Empirical stress factor models isolate the impacts of time, current, SoC,
temperature, and depth-of-discharge (DoD) on battery state-of-health (SoH). Through a few simplifying
assumptions, these stress factors can be represented using regularization norms. Physical degradation models
can further be classified into models of side-reactions and those of material fatigue. This article demonstrates
the importance of model selection to optimal control by providing several example controller designs.
Simpler models may overestimate or underestimate the capabilities of the battery system. Adding details
can improve accuracy at the expense of model complexity, and computation time. Our analysis identifies six
gaps: deficiency of real-world data in control literature, lack of understanding in how to balance modeling
detail with the number of representative cells, underdeveloped model uncertainty based risk-averse and
robust control of BESS, underdevelopment of nonlinear energy based SoC models, lack of hysteresis in
voltage models used for control, lack of entropy heating and cooling in thermal modeling, and deficiency
of knowledge in what combination of empirical degradation stress factors is most accurate. These gaps are
opportunities for future research.

INDEX TERMS Batteries, modeling, distributed energy resources, battery energy storage system (BESS),
state-of-charge (SoC), state-of-health (SoH), energy storage, optimal control.
Term Meaning
The associate editor coordinating the review of this manuscript and Ah - ampere-hour
approving it for publication was Khmaies Ouahada . BESS - battery energy storage system

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see http://creativecommons.org/licenses/by/4.0/
VOLUME 7, 2019 178357
 D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

BoL - beginning-of-life Simplified vector, vector equation, and vector function

BMS - battery management system notation
CRM - charge reservoir model - If z is a decision variable, z is a column vector of decision
EER - energy efficiency ratio variables. This is normally used to indicate decision variables
EMS - energy management system at discrete times (e.g., p at each time-step in a control horizon
EoL - end-of-life becomes p)
ERM - energy reservoir model - The expression z[1:3] denotes a column vector with the
EPRI - Electric Power Research Institute elements of z indicated by the index(es) (in this case, the first
HVAC - heating ventilation and air conditioning three elements)
- The expression z + y denotes a column vector that is the
kW - kilowatt
element-wise addition of the vectors z and y
kWh - kilowatt-hour
- For the scalar value b, the expression b[1] denotes the
LQR - linear quadratic regulator multiplication of the constant (b), times a vector of ones ([1]),
MPC - model predictive control that produces a column vector that is populated with b
P2D - pseudo two-dimensional model - The vector equation z + y = b[1] denotes a number of
SEI - solid electrolyte interphase equations equal to the length of z and y, each with indexed
SNL - Sandia National Laboratories variables (a.k.a, z[1] + y[1] = b, z[2] + y[2] = b, etc.)
SPM - single particle model - The vector equation z + y = b[1], can alternatively
SoC - state-of-charge be written as g(z, y) = [0] where g is the vector function
SoH - state-of-health g(z, y) = z + y − b[1]
SoL - state-of-life Parameters
ToU - time-of-use Model parameters are introduced and explained together
as there are too many to list here. All parameters used in the
Symbol Decision variable description application sections can be found in Tables 4, 5, 6, 11, and 13
p - ac power provided to the BESS respectively.
pdc - dc power provided to the battery
I. INTRODUCTION
ibat - dc current provided to the battery
Battery energy storage systems (BESS) can play an integral
vbat - battery terminal voltage
role in resilient and efficient power systems because of their
voc - open-circuit-voltage
ability to provide a range of energy services [1]. One of the
ς - state-of-charge
fundamental problems in BESS integration within the electric
cp - concentration in the cathode
power grid is designing control systems to maximize the
cn - concentration in the anode value of energy services provided [2]. BESS models used
8p - cathode half-cell potential in control systems formally represent assumptions about the
8n - anode half-cell potential physics underlying the conversion and storage of electrical
ηp - cathode overpotential energy. The BESS model is a critical element of effective
ηn - anode overpotential control and operation of BESS that, ultimately, enables more
τ - peak net electrical load resilient and efficient power systems.
T - battery surface temperature The control objective for a BESS often involves minimiz-
T0 - battery internal temperature ing an objective function (e.g., cost to the operator) subject
TEN - BESS enclosure temperature to the constraints of the system. The controller must decide
pHVAC - ac power provided to the HVAC unit of the settings for both real and reactive power (decision variables),
BESS within limits on power, energy, state-of-charge (SoC), volt-
%̇ - degradation rate (rate of change in SoH) age, current, temperature, and state-of-health (constraints).
fd - aggregate degradation stress factor Unlike in electric vehicles or consumer electronics (where
St - time degradation stress factor the controller is an element of the battery management sys-
Sς - state-of-charge degradation stress factor tem (BMS) [3]), the BESS controller is an element of the
ST - temperature degradation stress factor energy management system (EMS), which is responsible for
δ - depth-of-discharge issuing control decisions for all devices within its purview
(e.g., a home, building, microgrid, etc.). The BMS and EMS
Sδ - depth-of-discharge degradation stress factor
can share hardware and have overlapping responsibilities,
x - generalized decision variable in a given
as shown in Fig. 1, and often do not share a BESS model.
application problem (subscript denotes model
Models used in BMSs are often developed by the battery
used: ERM, CRM, SPM, T (temperature),
manufacturers themselves and hence can contain detailed
H (degradation))
information about underlying chemical process not available

178358 VOLUME 7, 2019

D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

detail. These points are different depending on application

and technology.
The goal of this article is to review the forms and functions
of BESS models with critical attention to their advantages,
disadvantages, and characteristics. This work will help read-
ers navigate the complex trade-offs involved in designing a
BESS controller. To explicitly define the scope of this review,
we start with a general optimal control problem and then
add detail relevant to BESS in order to classify different
aspects of battery models. A general optimal control problem
is formulated in:

FIGURE 1. Typical energy management system control diagram. min f (x)

x∈Rm
subject to: g(x) = [0]
h(x) ≤ [0] (1)
to an EMS controller. Further, for applications that require
accuracy at very high sample rates, such as power system sta- where x is a vector of decision variables, m is the total
bility, the inverter can have its own battery model for dynamic number of decision variables (roughly equal to number of
optimal control. This article focuses on optimal BESS control BESS model variables × the number of time steps in the
design within the EMS and so falls between the established optimization time horizon), f : Rm → R is the objective
fields of optimal control and battery modeling. function, g : Rm → Rr is a vector of equality constraints, and
The methods for designing optimal controllers for energy h : Rm → Rw is a vector of inequality constraints. In general,
storage systems have already been reviewed in [2]. Exam- for BESS applications, the objective function to be minimized
ple methods as applied to BESS include model predictive can be split into two terms: an objective associated with
control (MPC) [4]–[7], and linear quadratic regulator (LQR) battery operation and degradation (fb ) and an objective asso-
control [8], [9]. While there are large differences between the ciated with the service being provided (fs ). Further, the sets
methods for designing controllers, at the core of any approach of constraints can be split into constraints dealing with the
is a model of the battery system. In this article, we will largely service (e.g., peak load constraint) and constraints based on
ignore what method is used to design the controller, instead the equipment (e.g., maximum battery voltage). This split is
focusing on the commonalities and differences between the formalized in the multi-objective optimization problem in:
models. min fs (x) + fb (x)
BESS models mathematically represent the physical x∈Rm
dynamics and constraints of real systems. When choosing subject to: gs (x) = [0]
a BESS model, implicit assumptions are made about which gb (x) = [0]
physical dynamics are important to the controller’s operation hs (x) ≤ [0]
and which can be ignored. If a model ignores a state variable
hb (x) ≤ [0] (2)
(e.g., temperature) that ends up as a constraining factor in the
physical system, the control will be suboptimal. Similarly, The purpose of splitting the problem up is to isolate the
if a model inaccurately represents a system state variable components of the battery model (fb , gb , and hb ), as distinct
(e.g., battery state-of-charge), the controller will have to from those of the service model (fs , gs , and hs ).
constantly correct for the modeling error and again will be This article conducts a review of the battery model com-
suboptimal. However, consideration of which state variables ponents of the problem in (2). Specifically, the objective
to include and what physical dynamics to represent must functions fb (x) and constraints gb (x), and hb (x) associated
inevitably be balanced with the complexity of the model and with optimal control of BESS. The inequality constraints
the computational burden of the controller [10]–[12]. The hb (x) ≤ 0 ensure safe operation and battery longevity
chart in Fig. 2 conceptually illustrates the trade-off between (e.g., preventing over-temperature T ≤ Tmax , where T is the
model accuracy and complexity. The model categories to the battery temperature, and Tmax is the maximum temperature).
left on the chart are simple enough for control design and do The equality constraints generally represent a battery’s phys-
not require detailed knowledge of battery cell construction ical dynamics and the mathematical relationships between
and chemistry. At some level of model complexity there variables (e.g., pdc = vbat ibat , where pdc is the dc power,
is a tipping point where the improvements in accuracy are vbat is the battery voltage, and ibat is the battery current).
too costly, in terms of computation or level of information The expression fb (x) represents the BESS’s contribution
required, to be useful in control design. Further, as there are to the objective function based on the control action
thousands of individual cells in a BESS, there is logically (e.g., where the objective is to minimize costs, fb (x) may be
some point at which it is better to represent more cells at C EoL %̇, where C EoL is the end-of-life cost and %̇ is the rate
the same level of detail rather than increasing the level of of change in state-of-health (SoH)). Our goal is to present the

VOLUME 7, 2019 178359

 D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

FIGURE 2. Illustration of the trade-off between model accuracy and complexity (computational complexity equates to CPU time, and model
accuracy has also been referred to as ‘‘predictability’’ [12]).

advantages and disadvantages of various models to inform TABLE 1. Summary of case study assumptions.
design and further research on optimal control of BESS.
This article is organized as follows. Section II establishes
an example scenario used in each model domain section to
demonstrate its application. Section III introduces the various
models for state-of-charge, Section IV discusses temperature
models, and Section V discusses battery degradation models.
Each of these sections first introduces the model’s functions,
state variables and physical dynamics, and then includes a
representative controller design. Section VI discusses broad
trends and observations on the state-of-the-art including
identified gaps, and Section VII provides a summary and
conclusions.

II. PROBLEM STATEMENT

In this article we introduce models for different battery sys-
tem dynamics. To illustrate the impact of different classes of
models on control system performance, a representative con-
troller is formulated for each modeling domain. This section
establishes our example scenario in the form of a problem
statement. A summary of scenario assumptions can be found
in Table 1. FIGURE 3. Time-of-use price schedule (top), and customer electrical load
We consider a commercial electrical customer billed for (bottom) [15].
both time-of-use (ToU) energy and peak-demand charges.
This customer decides to purchase and install a battery to
reduce their electricity bill. The customer’s energy contract The load data used for this problem, as shown in Fig. 3 (bot-
charges 9 g/kWh during off-peak hours, 11 g/kWh during tom), are adapted from the EPRI test circuit ‘Ckt5’ loadshape,
partial-peak hours, and 15 g during peak hours according to normalized to a 1.0 MW peak [15]. We will assume that
the schedule in Fig. 3 (top) [13]. The utility then charges a the load and price are known a priori. Without the battery,
$50/kW service fee according to the peak net load measured the total bill would be calculated according to:
during the billing period. This price is consistent with demand
charges in specific localities in California and New York [14]. fs = 1t w> l + max(l) ν (3)

178360 VOLUME 7, 2019

D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

where l ∈ Rn is the load (kW) over time, w ∈ Rn is the where ς is the SoC, ςmax is the maximum SoC, and ςmin is
ToU energy price ($/kWh) over time, ν is the service fee in the minimum SoC. These box constraints are often enforced
$/kW for peak net load measured during the billing period, by a controller to ensure safety and design life, but many
and •> denotes a vector’s transpose. We use a time-step alternative methods for incorporating degradation into opti-
1t = 15 minutes (0.25 hours), and n = 96 (1 day). For mal control are discussed in Section V. Note that a BMS can
this problem we assume that the net-load is always greater also prevent overcharge/overdischarge by constraining SoC,
than zero. The total baseline electrical bill for this day is but these bounds are generally set at or outside the normal
$52,080 ($50,000 demand, $2,080 energy). With the addition operational range the controller uses.
of a BESS that can supply (−), or absorb (+), power p, Modeling SoC helps the controller know when in the future
the customer’s total bill can be modified to: it is likely to encounter these limits and to make control deci-
fs (p) = 1t w> (l + p) + max(l + p) ν (4) sions accordingly. In optimal control, SoC models inform the
controller how control decisions affect future SoC and enable
where p is the battery system power that element wise sub- the controller to adjust decisions to optimize an objective.
tracts from l when the battery system is discharging. The Therefore, errors in SoC models can lead to poor control
problem formulation can be expressed as: design an opti- performance or even infeasible solutions.
mal battery dispatch control scheme that minimizes the cus- The various models for SoC can be classified by the
tomer’s total bill subject to the constraints of the battery and units with which they define nominal and available capacity
the customer’s system. The dispatch is open-loop, and we according to (5). Models that define capacity in units of
do not consider modeling uncertainty in this control scheme. energy (kWh) can be classified as energy reservoir models
Hence, we do not consider the mismatch between the con- (ERMs), those which define it in units of charge (Ah) can
troller model and a real system. Research into the effects be classified as charge reservoir models (CRMs), and those
of modeling uncertainty on BESS controller performance is which define it in units of concentration (mol/L) can be
ongoing [16]. classified as concentration-based models. ERMs, discussed
in Section III-A, do not include dc voltage or current, so they
generally have fewer variables and constraints. This simpli-
fication can lead to unrepresented physical dynamics that
can, under some circumstances, have negative effects. CRMs,
discussed in Section III-B, include expressions to represent
current-voltage (I-V) dynamics which can improve accuracy
at the expense of increased model complexity. Concentra-
tion based models, discussed in Section III-C, include many
parameters associated with the specific electrochemical reac-
FIGURE 4. Thought experiment demonstrating how the electrochemical tion and cell design that can predict battery dynamics. Each
definition of the SoC of a battery loses physical meaning when applied to
strings [17].
has its appropriate applications in control design as well as
difficulties and drawbacks. Table 2 shows a summary of these
trade-offs as discussed in detail in the following subsections.
III. STATE-OF-CHARGE MODELS
Electrochemically, a battery cell’s SoC is related to the con- A. ENERGY RESERVOIR MODELS
centration of the limiting active species, in the relevant reac- ERMs are a class of SoC model that define capacity in units
tion at the associated electrode [17]. This physical association of energy (kWh). An example ERM for SoC is shown in:
however, breaks down when the electrochemical definition of
SoC is applied to strings as the thought experiment in Fig. 4 ∂ς
illustrates. When referring to BESS, it is more common to use Qcap = ηe p+ + p− (7)
∂t
an empirical definition of SoC, represented in:
Available Capacity where ς is the SoC, p+ and p− are the charge and discharge
State of Charge , (5) ac power respectively, Qcap is the energy capacity, ηe is the
Nominal Capacity
round trip energy efficiency, and ∂ς/∂t represents the rate
which is the ratio of available to nominal capacity. Normal- of change of SoC. To make this constraint convex, charge
izing SoC to the range [0,1] or [0%,100%] is an intuitive power and discharge power are formulated as independent
simplification, especially as nominal capacity can change decision variables. While this means that simultaneous charge
over time, but it is not mathematically necessary to do so. and discharge would not violate the explicit constraints,
In this context, capacity can be measured in energy with units the objective function is often structured such that there is no
of kilowatt-hours (kWh), charge with units of ampere-hours advantage to candidate solutions that do so. Hence, as long as
(Ah) or in concentration with units of moles-per-liter (mol/L). energy prices are positive, and efficiency is in the range [0, 1],
Constraints on SoC are shown in: the optimal solution to a control problem with this SoC con-
ςmin ≤ ς ≤ ςmax (6) straint will always satisfy complementary slackness between

VOLUME 7, 2019 178361

 D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

TABLE 2. Summary of SoC models.

charge and discharge power. When objective does not have that is based on many assumptions, so it may or may not be
these properties, an additional non-convex constraint can be appropriate for a given application.
added to prevent solutions with simultaneous charge and Some previous work using ERM ignore efficiency losses
discharge (p+ p− = 0). entirely [7], [18], [20]–[22]. However, due to the error it
Rather than a constant, as shown in (7), the energy effi- incurs, this is ill-advised for controllers that schedule SoC
ciency of a BESS can be a time-varying, nonlinear function over any significant time horizon. Much work includes both
of battery SoC, voltage, current, temperature, and state-of- charge and discharge efficiencies [6], [23]–[25], [27]–[32].
health (SoH). Assuming a constant energy efficiency can, Self-discharge power can also be included in an ERM
by extension, be an implicit assumption these states are also [6], [25], [27], [31]–[34] as shown in:
constant. Some of these assumptions are valid for a range
∂ς
of applications. SoH, for instance, changes very slowly with Qcap = ηe p+ + p− + psd (8)
respect to a control horizon. Other assumptions however, are ∂t
only valid for a narrow operational range. How wide the oper- where psd is the self-discharge power.
ational range can be, while the ERM remains a sufficiently We refer to (8) as a Type 1 model, in that it only includes
accurate approximation, depends on how flat the energy charge efficiency. Models that include both charge and
efficiency curve is with respect to each variable. For example, discharge efficiencies are referred to as Type 2 models, while
changing battery voltage can change BESS efficiency but those that only include discharge efficiency are referred to
some battery types have a wide range of SoC where the open- as Type 3 models. These model types are able to produce
circuit-voltage is nearly constant. ERMs are more accurate equivalent relationships between power and the rate of change
over a wider range of SoC for these types of batteries than in SoC over time. Table 3 shows the conversion calculations
for a battery whose open-circuit-voltage changes quickly with needed to move from one type to another while maintaining
respect to SoC. ERMs use a simple representation of SoC this equivalence. We use a Type 1 model in this paper (with

178362 VOLUME 7, 2019

D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

TABLE 3. Conversion between equivalent efficiency representations.

FIGURE 5. Energy Based Kinetic Battery Model [36].

FIGURE 6. Feasible region for ac power described by the kinetic battery

model.

kinetic battery model splits the reservoir into available energy

and bound energy as shown in Fig. 5 and in:
∂ς1
(1 − cf )Qcap = 0e (ς2 − ς1 ) (9a)
∂t
∂ς2
ηcha = ηe the round trip energy-efficiency) because it has cf Qcap = ηe p+ + p− + 0e (ς1 − ς2 ) (9b)
∂t
the intuitive property that the SoC multiplied by the capacity
directly reflects how much energy is expected to be available where cf is the fraction of total capacity in the available
from the battery on discharge. reservoir, and 0e is a time constant that governs the rate
While most ERM use ac power, a few use dc power [59] by of energy transfer between the two reservoirs. The physical
including many of the constraints in the CRM. The manufac- intuition of this model is that the higher the discharge rate the
turer dc kWh rating for batteries is calculated based on either more quickly the available energy is depleted, and the low-
the Ah rating multiplied by the nominal battery voltage or the level limit of the tank is reached. This is equivalent to the
energy extracted during a constant current discharge test. linear inequality constraints on power in:
Hence, this approach is still subject to the inaccuracy of other pmin ≤ p ≤ pmax (10a)
ERM, over a wide operational range, if voltage is far from m1 ς + b1 ≤ p ≤ m2 ς + b2 (10b)
the nominal voltage assumed or the constant current rate used
during testing. where pmin is the discharge power limit, pmax is the charge
Though it is not commonly done, all three parameters can power limit, m1 and b1 are the slope and intercept of the
be functions of the SoC or temperature or both (Qcap (ς, T ), linear power limit on discharge, respectively, and m2 and
ηe (ς, T ), and psd (ς, T )) [38]–[40]. The most common version b2 are the slope and intercept of the linear power limit on
of this is SoC dependent losses psd (ς) = msd ς + bsd where charge, respectively. These constraints are encountered at
msd is the proportional power loss and bsd is the power high discharge rate more quickly according to the slope m1
loss at ς = 0 [23], [26]. Adding nonlinearity to these and intercept b1 . These constraints then reduce maximum
functions has the potential to increase predictive accuracy power linearly as SoC approaches its minimum, the same as if
over a wider operational range of SoC and warrants further the maximum power is constrained by the difference between
investigation [38]. tank levels in the kinetic battery model. The limits defined in
To represent the relationship between SoC and the power (10) are shown in Fig. 6
limits (pmax and pmin ) a two reservoir ERM, also called the Injection and absorption of reactive power can be an impor-
kinetic battery model, is sometimes used [36], [37]. The tant capability for BESS in many applications [1]. While

VOLUME 7, 2019 178363

 D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

reactive power does not directly affect SoC, it can indirectly time step, and the matrix D is defined below.
affect SoC by constraining real power. Constraints on appar-
0 . .
 
−1 1 0
ent power and power factor are presented in [28]:  0 −1 1 0 . .
1 
. .

p = p+ + p− (11a) D=   (13)
1t 

. .


p2 + q2 ≤ Smax
2
(11b)
0 0 −1 1 n×(n+1)
p
p ≥ p.f.min (11c) The SoC constraint in (12) can be included with the kinetic
p2 + q2
battery model constraints in (10) yielding the problem formu-
where q is the reactive power (var), Smax is the apparent power lation in:
limit, and p.f.min is the minimum power factor. While there min 1tw> (l + p+ + p− ) + ντ (14a)
are no direct incentives for reactive power, there might be x ERM ∈R3n+2
penalties on poor power factors. Therefore in many cases, subject to: Qcap Dς = ηe p+ + p− + psd [1]
constraint 11 must be enforced.
ς [1] = ς0 (14b)
TABLE 4. Energy reservoir model parameters.
ς [1] = ς [n] (14c)
pmin [1] ≤ p+ + p− ≤ pmax [1] (14d)
ςmin [1] ≤ ς ≤ ςmax [1] (14e)
m1 ς + b1 [1] ≤ p + p ≤ m2 ς + b2 [1]
+ −

(14f)
l + p + p ≤ τ [1]
+ −
(14g)

where x ERM = p , p , ς, τ ∈ R3n+2 , and τ ∈ R is

 + −

a dummy variable that represents the peak net load. The

constraint (14b) ensures that control decisions are made based
on the current estimated SoC. The constraint (14c) represents
the intuitive assumption that the BESS will continue to oper-
ate after the end of the current control horizon and that the
next period will be similar to this one. In this application,
(14c) is used to make simulation results easier to interpret
and compare. The objective has been modified to use the
dummy variable τ to represent peak load in the objective and
add a constraint that it be greater than the net load at every
time (14g).
1) ERM APPLICATION In the code accompanying this article the minimum heat
In this section we solve the problem outlined in Section II generation regularization term, described in Section V-A.1
with an optimal controller designed using an ERM. The ERM (5||p+ + p− ||22 , with a very small weight 5 = 1e-5),
is used here to demonstrate its application, however, it is is applied to the objective in this and each application script.
not the most appropriate model for this problem because of This has the effect of avoiding spikes or abrupt changes in
its inaccuracy over a wide range of voltage in this scenario. power, while not significantly impacting the minimum value
In cases where the performance of the ERM model is less than achieved.
desirable, it is sometimes employed because of the computa- Information on numerical algorithms for solving general
tional simplicity. The example scenario listed here serves to linear and nonlinear optimization problems can be found
demonstrate how the ERM can be applied to solve a simple in [60], [61]. We solve this using the Pyomo optimization
problem, and one can extrapolate it to how it could be used to modeling language [62], [63] and the Ipopt interior point
solve a more complicated problem. Example parameters for optimization problem solver [64]. The solution is shown
the ERM are listed in Table 4. in Fig. 7. Note that even though the maximum discharge
We can express the constraint described in (8) between power is 500 kW, the battery is only able to reduce the peak
each SoC using the vector equation (12). net load by approximately 85 kW because of limitations
on energy. The effect of the kinetic battery model can be
Qcap Dς = ηe p+ + p− + psd [1] (12) observed around hour 10, when the battery finishes charging
then pauses for one time step only to then charge at a low
where ς ∈ Rn+1 is the SoC at each time step, p+ ∈ Rn+ and level to maintain 94.8% SoC. This artifact of the model is
p− ∈ Rn− are the ac charge and discharge power during each because the maximum charge rate at 95% SoC is 0 kW which

178364 VOLUME 7, 2019

D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

parameters of Peukert’s equation assume a constant-current

discharge, where 0% SoC corresponds to battery reaching its
minimum voltage under load. This model implies a battery
equivalent circuit, that is better to represent explicitly, and a
static operating condition (constant-current discharge) that is
one of the decision variables in our problem formulation.
While some previous work ignore efficiency losses
[41]–[45], this many not be accurate for controllers that pre-
dict SoC over an extended time horizon for the same reason
as discussed in Section III-A. Self-discharge current can be
included in a CRM [42], [52], [68] as in:
∂ς
Ccap = ηc i+ −
bat + ibat + isd (17)
∂t
where isd is the self-discharge current. While less common
for CRM, efficiency in this model can be represented equiv-
alently with Type 2 and Type 3 models as shown in Table 3.
FIGURE 7. Results calculated with the ERM: (a) Net load with BESS power
control, (b) Battery power, (c) Battery SoC (bottom).

is insufficient to counter self-discharge power. The ERM

expects to be able to reduce the peak to 914.7 kW and it
is clear from the calculated net load the schedule allocates
charging to the periods of low electricity price.
The control solution reduces the total electrical bill
from $52,080 ($50,000 demand, $2,080 energy) to $47,837
($45,737 demand, $2101 energy). The net effect is a
$4,243, or 8.15%, reduction in the electrical bill. Note that
while the demand charge is reduced significantly, the energy
bill increases due to efficiency losses in the BESS. FIGURE 8. Equivalent Circuit Models.

B. CHARGE RESERVOIR MODEL (CRM) Several additional constraints are needed to govern inter-
CRMs are a class of BESS models that define capacity in nal relationships between voltage, current, dc power, and ac
units of charge (Ah). An example CRM is shown in: power. The foundation of these is an equivalent circuit model
∂ς [49], [50], [69], [70]. The most common battery equivalent
Ccap = ηc i+ −
bat + ibat (15) circuit models are shown in Fig. 8 and described in:
∂t
ibat = i+ −
bat + ibat (18a)
where i+ −
bat and ibat are the charge and discharge current respec-
tively, ς is the battery SoC, Ccap is the charge capacity, ηc ∂v1 −1 1
= v1 + ibat (18b)
is the coulombic efficiency, and ∂ς/∂t represents the rate of ∂t R1 C1 C1
change of SoC. Like with the ERM, to make this constraint ∂v2 −1 1
= v2 + ibat (18c)
convex, charge current and discharge current are formulated ∂t R2 C2 C2
as independent decision variables. Simultaneous charge and voc + R0 ibat + v1 + v2 = vbat (18d)
discharge is avoided in the same way, by structuring the where R0 , R1 , C1 , R2 , and C2 are equivalent circuit resistor
objective function such that there is no advantage to those and capacitor parameters. The 0th order equivalent circuit
candidate solutions. is accurate for steady state analysis as it accounts for bat-
Peukert’s equation relates the charge capacity to the dis- tery ohmic resistance R0 but not any time-domain dynamic
charge rate in amps [65], [66]: response. The 1st and 2nd order models are increasingly
Ccap = (i− kpeu accurate for analyses requiring short time steps (roughly
bat ) tdischarge (16)
faster than 10 minutes between samples or (1/600)Hz) [48].
where kpeu is the Peukert exponent, and tpeu discharge The R-C parallel elements of the circuit can represent
time before the battery reaches its low voltage limit. Peuk- different chemical reaction dynamics within battery cells:
ert’s equation is sometimes used in control design [67]. R1 & C1 can represent ion-diffusion (Warburg impedance)
However, it makes several simplifying assumptions that whereas R2 & C2 can represent anode-cathode capaci-
do not make sense for optimal control applications. The tance or constant phase element [69]. Note that the time

VOLUME 7, 2019 178365

 D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

voc = kT (T − Tref )(mς + v0 ) (21b)

voc = aς 2 + bς + c (21c)
voc = ας 3 + βς 2 + γ ς + δ (21d)
(1 − ς)
voc = bk − mk + ck edk ς (21e)
ς
voc = α(ς)ς 3 + β(ς)ς 2 + γ (ς )ς + δ(ς) (21f)

FIGURE 9. Charge Based Kinetic Battery Model [72], [73]. where vm and v0 are the slope and intercept of a linear voc
model, respectively, kT is a linear temperature adjustment,
T and Tref are the battery temperature and reference battery
constant of the diffusion element (τ1 = R1 C1 ) is generally temperature, respectively, a, b, and c are the coefficients for
much larger than the time constant of the capacitance element a quadratic polynomial fit, α, β, γ , and δ are the coefficients
(τ2 = R2 C2 ). The equivalent circuit’s impedance parame- for a cubic polynomial fit, bk , mk , ck , and dk are the coeffi-
ters can be calculated using least squares system identifica- cients for a negative reciprocal and exponential function fit,
tion [71]. These parameters can be functions of current, SoC, and α(ς), β(ς), γ (ς), and δ(ς) are piecewise functions that
SoH, temperature, or any combination thereof [52]. collectively comprise a cubic spline. The simplest function
There are many variations of these equivalent circuits. for voc is a linear approximation (21a) which can be accurate
Adding a resistor across the voltage source is equivalent to within a narrow range of SoC [4], [35], [50], [74]. A temper-
making isd a linear function of voc . Adding a resistor across ature adjustment can also be applied (21b) to improve accu-
the battery terminals is equivalent to making isd a linear func- racy [50], [68]. Polynomial approximations are also used,
tion of vbat . Several other configurations are discussed in [48], (21c) or (21d), but these are sometimes non-convex and so
but it is unclear how these additions affect the accuracy of the can be more computationally intensive to work with. Another
model. approach is to model voc as a combination of a negative recip-
Though it is rare to do so in a controller, battery voltage rocal and exponential functions (21e) [75], [76]. This model
hysteresis can be incorporated into the equivalent circuit works better for lithium-cobalt batteries or other chemistries
model [17] as shown in: with exponential curves near 100% and 0% SoC but that
∂vhys are relatively flat and straight in a wide range around 50%
= γhys sgn(ibat ) (M (ς, ibat ) − vhys ) (19) SoC. Note that (21e) has an asymptote at ς = 0, and hence
∂t
the model must constrain SoC to some positive threshold to
where vhys is the dynamic voltage hysteresis, γhys is a decay
work well. Piecewise cubic splines (21f) are the most accurate
rate tuning constant, and M : R2 7 → R is a function
[48], but these can be very difficult functions to work with in
that returns the maximum voltage hysteresis. Specifically,
optimization. Example open-circuit-voltage data along with
M (ς, ibat ) is an empirical approximation based on experimen-
different fit types are shown in Fig. 10. Battery voltage
tal data that is positive for charge and negative for discharge.
hysteresis can alternatively be represented within the open-
Alternatively, the hysteresis can be modeled using an addi-
circuit-voltage function by modeling voc differently on charge
tional charge reservoir as in [72], [73]. This approach splits
and discharge [75].
the total charge capacity into two states: bound charge, and
Battery power is modeled through Ohm’s power law:
available charge as shown in Fig. 9 and in:
∂ς1 pdc = ibat vbat (22)
(1 − cf )Ccap = 0c (ς2 − ς1 ) (20a)
∂t where pdc is the dc power.
∂ς2
cf Ccap = ηc i+
bat + ibat + 0c (ς1 − ς2 )
−
(20b) The conversion efficiency from ac to dc power, or vice
∂t versa, is sometimes ignored. When it is modeled, conversion
where cf is the fraction of total capacity in the available efficiency is commonly modeled as a constant [28]. One way
reservoir, and 0c is a time constant that governs the rate of modeling inverter efficiency as a constant is shown in:
of charge transfer between the two reservoirs. The open-
1 −
circuit-voltage is then based on the available charge level pdc = φe p+ + p (23)
only. This effectively represents energy recovery effect and φe
is structurally similar to the discrete version of the single where φe ∈ [0, 1] is the conversion efficiency con-
particle model discussed in Section III-C. stant. Alternatively, a linear fit or quadratic fit can be
Open-circuit-voltage voc , also referred to as electromotive used [77]:
potential or force, is the terminal voltage of the battery when
measured ‘at-rest’ and is a function of the SoC, SoH and pdc = φm p + φb (24)
temperature of the cell. Several example functions for voc are pdc = φ0 p2 + φ1 p + φ2 (25)
given in:
where φm and φb are the slope and intercept of a linear
voc = vm ς + v0 (21a) efficiency function, respectively, and φ0 , φ1 , and φ2 are the

178366 VOLUME 7, 2019

D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

‘‘Short board effect’’ method: This method uses models

of the extreme cells in a string to better represent limiting
factors. The maximum SoC cell and minimum SoC cell are
tracked independently, each with their own decision variables
and limits. The total string voltage is then the sum of the
highest SoC cell voltage, the lowest SoC cell voltage, and
the voltage of a ‘‘big cell’’ representing all remaining cells.
This approach more accurately models when the string will
encounter cell voltage or SoC limits. There is some increase
in the computational complexity as there must be additional
decision variable for voltage, dynamic voltage, and SoC as
well as their associated constraint sets.
‘‘One-by-one calculation’’ method: This method explic-
itly represents all cells in a battery pack. In cases where cells
have a wide variance in capacity and coulombic efficiency,
it is possible that the highest SoC cell and/or lowest SoC
cell will switch cells within the control horizon. This method
will be able to predict and optimize operation whichever cell
is the limiting factor for a given cycle. Explicitly modeling
every cell within a string also enables the direct represen-
tation of cell balancing circuits within the controller, as in
FIGURE 10. Open-Circuit-Voltage Models. [79], [80]. This also enables the representation of unequal cur-
rent splitting in parallel cells or strings. With hundreds or even
thousands of cells in a grid scale BESS, this approach
coefficients of a quadratic efficiency function. Inverter effi- can easily become computationally infeasible to apply in
ciency can be a nonlinear function of ac voltage, dc voltage, practice.
and temperature [78]. As CRMs account for changes in battery dynamics over
While CRMs normally include box constraints on SoC and the range of voltage and current, they are more accurate than
real/reactive power, additional box constraints on current and ERMs in applications where the SoC and charge/discharge
battery voltage are shown in: currents vary significantly. However, given the increase in
imin ≤ ibat ≤ imax (26a) complexity, they are much more difficult to use in the design
of optimal controllers. Hence, CRMs are best used in appli-
vmin ≤ vbat ≤ vmax (26b)
cations relying on long duration, if sparse, charge/discharge
Unlike the ERM, it is not necessary to add SoC dependent schedules (e.g., day-ahead hourly energy arbitrage). In such
power constraints as the voltage constraints handle these applications the high rate battery dynamics in (18) can often
limits implicitly. be ignored.
An important factor to consider is that battery cells within
a string may have significantly different parameters resulting 1) CRM APPLICATION
from normal manufacturing variation. When using a CRM In this section we solve the problem outlined in Section II
in a controller design, there are at least three methods for with an optimal controller designed around a CRM. Exam-
accounting for distributions in parameters and states within ple parameters for a CRM are listed in Table 5.
a BESS [69], [79]. What follows is a discussion of these The ‘CRM: no dynamic voltages’ from Table 2 is the
methods and their relative advantages for optimal control. most appropriate model for this problem because of the
‘‘Big cell’’ method: This method is based on a simplifica- long forecast horizon, low (15 minute) time resolution,
tion that models a battery pack as one large battery cell. In this and because we are only controlling one battery system.
approach, the battery voltage is the individual cell voltage We also use the ‘‘big battery’’ approach to modeling the
multiplied by the number of cells in series, the capacity dc battery string. Implementing the CRM into a usable for-
is the total capacity and so on. Intercell balancing is gen- mat requires reformulating the differential equation for SoC
erally handled within the self-discharge current parameter. defined in (15) into the vector of equality constraints as
This method is most accurate when a string is made up of shown in:
very well-matched cells, which are manufactured to have a
very narrow distribution of performance. For poorly matched Ccap Dς = ηc i+ −
bat + ibat (27)
cells, or for cells that have degraded and hence have widened
in their performance distribution, this approach is optimistic where ς ∈ Rn+1 is the vector of SoC at each time step,
in its approximation of string capacity on both charge and i+ n − n
bat ∈ R+ and ibat ∈ R− are the vectors of charge and
discharge. discharge dc current respectively, and D is a matrix defined

VOLUME 7, 2019 178367

 D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

TABLE 5. Battery system charge reservoir model parameters. assumption that the BESS will continue to operate after the
end of the current control horizon and that the next period
will be similar to this one. In this application, (28h) is used
to make simulation results easier to interpret and compare.
Again, while simultaneous charge and discharge does not vio-
late explicit constants, the structure of the objective ensures
that solutions will comply with complementary slackness
between charge and discharge current.

above in (13). Like with the ERM, charge current and dis-
charge current are formulated as separate decision variables.
While this does not make the optimization problem convex,
it does improve the convergence time of the solver without
affecting the solution. The resulting problem formulation is
shown in:
min 1t w> (l + p) + ντ (28a)
x CRM ∈R6n+2
pdc = φ0 p2 + φ1 p + φ2 (28b)
pdc = (i+ −
bat + ibat )vbat (28c)
vbat = voc[1:n] + R0 (i+ −
bat + ibat ) (28d)
voc = ας + βς + γ ς + δ
3 2
(28e)
Ccap Dς = ηc i+ −
bat + ibat (28f)
ς [1] = ς0 (28g) FIGURE 11. Results calculated with the CRM: (a) net load with BESS
power control, (b) battery power, (c) battery SoC, (d) battery current, and
ς [1] = ς [n] (28h) (e) battery voltage.
pmin [1] ≤ p ≤ pmax [1] (28i)
ςmin [1] ≤ ς ≤ ςmax [1] (28j) The resulting customer’s net load and optimal control
schedule for the BESS are shown in Fig. 11 (a) and (b)
vmin [1] ≤ vbat ≤ vmax [1] (28k) respectively. The simulated battery current and voltage
imin [1] ≤ i−bat ≤ [0] (28l) are shown in Fig. 11 (c) and (e) respectively. The CRM
[0] ≤ i+
bat ≤ imax [1] (28m) based controller expects to be able to reduce the peak
load by approximately 83 kW. The peak battery voltage
l + p ≤ τ [1] (28n)
reached 780 V and the dc current reached −120 A on dis-
where x CRM = {p, pdc , ibat , vbat , voc , ς , τ } ∈ R6n+2 , pdc ∈ charge. The control solution reduces the total electrical bill
Rn is the dc electrical power provided to the battery, vbat ∈ Rn from $52,080 ($50,000 demand, $2,080 energy) to $47,948
is the battery terminal voltage, voc ∈ Rn+1 is the battery open- ($45,871 demand, $2077 energy). The net effect is a $4,132,
circuit-voltage, and τ is the dummy variable for peak power. or 7.93%, reduction in the electrical bill.
The CRM includes constraints on inverter conversion effi- If we assume that the example ERM and CRM represent
ciency (28b), Ohm’s law relating dc power, voltage and cur- the same physical BESS, then we can investigate which one
rent (28c), the battery equivalent circuit model (28d), and is a better controller. We can observe that the ERM expects to
the open-circuit-voltage curve (28e). The constraint (28g) be able to reduce the peak load, and the total bill, more than
ensures that control decisions are made based on the current the CRM. If the ERM is the more accurate model, then the
estimated SoC. The constraint (28h) represents the intuitive CRM will underutilize the batteries. However, if the CRM is

178368 VOLUME 7, 2019

D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

xs,n,ave − xs,n,0%
more accurate, as we expect it to be, then the control solution ς = (31)
from the ERM controller may be infeasible. xs,n,100% − xs,n,0%
xs,p,ave − xs,p,0%
= 1− (32)
C. CONCENTRATION BASED MODELS xs,p,100% − xs,p,0%
Concentration based models measure capacity in units of the where ς is the SoC of the cell, cs,j,max is the maximum
concentration (mol/L) of the active materials of the elec- concentrations of electrode j, and r̄j is the radius of the
trodes. These models can be further classified into single par- representative particle.
ticle model (SPM), pseudo-two-dimensional model (P2D), Battery voltage in the SPM is based on the open cir-
and many others [12]. In this section we briefly cover SPM cuit voltage, the chemical overpotential, and the electrical
and P2D models as these are the concentration models most resistance. while the SoC is based on average concentration
widely used in controllers. Note that concentration mod- throughout the particle, open-circuit-voltage is based only
els may require parameters based on cell construction and on its surface concentration [17]. These relationships are
chemistry that manufacturers consider proprietary and would shown in:
not be available to the controller. However, there are meth-
voc = 8p xs,p,surf − 8n xs,n,surf



ods available to estimate some or all of these parameters (33)
empirically [55], [58]. cs,p |r=r̄p
xs,p,surf = (34)
cs,p,max
cs,n |r=r̄n
xs,n,surf = (35)
cs,n,max
where 8p : [0, 1] 7→ R and 8n : [0, 1] 7→ R are
the positive and negative electrode potentials as functions of
their normalized surface concentrations (xs,p,surf and xs,n,surf
respectively). Like with open circuit voltage in the CRM, 8p
and 8n can be approximated using polynomial or exponential
functions. Authors in [17], [53] use a Redlich-Kister expan-
sion as a general best fit function for 8p and 8n :
 
FIGURE 12. Single particle model (SPM). RT 1 − xs,j,surf
8j = v0bat + ln
F xs,j,surf
SPMs represent each electrode as a single particle ( N
X Ak 
[54], [81]–[83] which is useful for modeling the effects of + (2xs,j,surf − 1)k+1
transport phenomena but loses some accuracy at high cur- F
k=0
rent, or wherever variations across the electrodes are signif-

2xs,j,surf k(1 − xs,j,surf )
icant [11], [84]. Figure 12 shows an simple generic SPM. − (36)
(2xs,j,surf − 1)1−k
The differential equation for mass balance in an intercalation
particle is governed by Fick’s law in a spherical coordinate where 8j is the potential at electrode j, xs,j,surf is the nor-
system [85], [86]: malized surface concentration at electrode j, R is the ideal
gas constant (8.314 J mol/K), T is the battery temperature
∂cs,j Ds,j ∂ 2 ∂cs,j
 
= 2 rj (29) in Kelvin,1 F is Faraday’s constant 96,487 coulombs/mol,
∂t rj ∂rj ∂rj k is the summation index number, and N , v0bat and Ak are
the fitting parameters. Figure 13 shows example anode and
where cs,j is the concentration of electrode j as a function
cathode equilibrium potential functions. When fully charged,
of both time t and particle radius rj , Ds,j is the solid phase
the active material concentration is at its maximum in the
diffusion coefficient, and the subscript j ∈ {p, n} represents
anode and at its minimum in the cathode. This means there
the positive/negative electrode. The SoC is a function of
is potential for ion movement from anode to cathode and
the average normalized concentration (stoichiometry) in each
electron movement from cathode to anode (a.k.a. discharge).
electrode (30). To calculate SoC, we first define xs,j,100%
Chemical overpotential can be calculated according to the
and xs,j,0% as the stoichiometry at which electrode j is at
Butler-Volmer equation [17], [87], [88]:
its maximum and minimum respectively. Using these defini-
tions, SoC is the state of the anode’s stoichiometry between ibat 1−αc αc
Jj = = kj cs,j,max c1−α
e
c (1 − x
s,j,surf ) xs,j,surf
xs,n,100% and xs,n,0% (or equivalently, 1 - the state of the as AL
cathode’s stoichiometry between xs,p,100% and xs,p,0% ) [17] (1 − αc )F αc F
    
as expressed in: × exp ηj − exp − ηj (37)
RT RT
Z r̄j
1 1 Note that battery temperature can be assumed to be constant, or this can
xs,j,ave = cs,j dr (30)
r̄j cs,j,max 0 be coupled with one of the thermal models discussed in Section IV

VOLUME 7, 2019 178369

 D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

FIGURE 14. Pseudo two-dimensional model (P2D).

1) SPM APPLICATION
FIGURE 13. Equilibrium potentials (open-circuit-voltages) of In this section we solve the problem outlined in Section II
lithium-cobalt-oxide (LiCO2 ) cathode (top) and MesoCarbon MicroBeads
(MCMB) anode (bottom) [53].
with an optimal controller designed around a SPM.
Example parameters for the SPM are listed in Table 6 with
Redlich-Kister expansion parameters for the anode and cath-
  
Ek,j 1 1 ode voltages listed in Table 7. The SPM is used here to
kj = kj,ref exp − (38) demonstrate its application, however, it is not the most appro-
R T Tref
priate model for this problem because of its high complexity
where ηj is the reaction overpotential, Jj is the current density relative to the time resolution and scale required. This model
on the particle’s surface, as is the specific interfacial sur- would be more appropriate for higher sample rate appli-
face area (volumetric fraction of the active material ×3/r̄j ), cations where voltage dynamics are more salient. Further,
A is the current collector area, L is the electrode thickness, kj the model parameters used here are derived from literature
is the Arrhenius rate of the electrochemical reaction, Ek,j is sources on cell-level design. With calculated capacity of
the activation energy of the Arrhenius relationship, ce is the roughly 1.9 Ah per cell, 445 parallel cells were simulated to
concentration of the electrolyte, and αc is the charge-transfer achieve a comparable capacity to the CRM (800 Ah). This
coefficient. This equation can be solved for ηj in terms of means that the ‘‘Big Cell’’ modeling assumption extrapolates
ibat making it possible to compute the Jacobian metrics with the performance of a single cell to 445×196 = 87, 220 cells.
respect to the parameters as was demonstrated in [88]. One advantage of the SPM is that it enables investigation
Electrical resistance is a combination of resistances in the of how changes to cell level design parameters might affect
electrolyte, the current collectors, the tabs, and the terminals. simulated system level performance.
These can all be modeled using an single constant resistor Within each particle we model five discrete volumes to
Rcell , but it is also common to apply a temperature correction approximate the radial dimension of the model. The core
factor, current correction factor, or both [88]. With the open- volume is spherical with radius drj which is surrounded by
circuit-voltage, the chemical overpotential, and the electrical four shell volumes each with a thinness of drj . Fig. 15 illus-
resistance calculated, the SPM battery voltage is shown in: trates how electrical current is transformed to current density
which is transformed in turn to changes in the chemical
vbat = 8p − 8n + ηp − ηn + Rcell ibat (39) concentrations within the modeled volumes. Fick’s second
law describes a concentration gradient in the representative
To make this model more accurate at high currents, we can particle that can be approximated using discrete volumes,
extend it to an additional spatial dimension along the length each being shells around a spherical core. Each of these
from the anode current collector, through the separator, to the shells has a chemical concentration capacity proportional
cathode current collector, as illustrated in Fig. 13. With to its volume and the maximum concentration. The surface
one dimension along the cell’s thickness and the pseudo between each volume has a chemical resistance proportional
dimension describing a concentration gradient within spher- to the surface area and inversely proportional to the diffusion
ical particles, this is called a pseudo-two-dimensional (P2D) coefficient. The current density at each particle’s surface
model [17], [89], [90]. Whereas with the SPM, ce , xs,j,surf , is proportional to the battery current (ibat ). The resulting
ηj , and therefore the Jj are essentially averaged over each optimization problem is formulated in:
electrode, the P2D represents these quantities as functions of
min 1t w> (l + p) + ντ
the dimension from one current-collector to the other [91]. x SPM ∈R18n+13
Full order P2D built with the partial differential equations subject to: (40a)
are too computationally complex for most real-time control
pdc = φ0 p2 + φ1 p + φ2 (40b)
applications [57]. However, discretized or reformulated P2D
models can be applied successfully in control applications pdc = ibat vbat (40c)
[10], [57], [58]. vbat = 8p[1:n] − 8n[1:n] + ηp − ηn + R0 ibat

178370 VOLUME 7, 2019

D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

FIGURE 15. Concentration reservoir analogy of the SPM with five discrete volumes per particle.

TABLE 6. Battery System Single Particle Model (SPM) parameters. TABLE 7. Equilibrium potential Redlich-Kister expansion parameters,
reproduced from [53].

    
0.5 F 0.5 F
× exp η − exp − η (40f)
RT j RT j
Sj,0 ibat Ds,j Sj,1 (cj,1[1:n] −cj,0[1:n] )
Vj,0 Dcj,0 = −
F as,j As,j Ls,j dr
(40g)
dr Vj,1
Dcj,1 = Sj,1 (cj,0[1:n] − cj,1[1:n] )
Ds,j
+ Sj,2 (cj,2[1:n] − cj,1[1:n] ) (40h)
dr Vj,2
Dcj,2 = Sj,2 (cj,1[1:n] − cj,2[1:n] ) (40i)
Ds,j
+ Sj,3 (cj,3[1:n] − cj,2[1:n] ) (40j)
dr Vj,3
Dcj,3 = Sj,3 (cj,2[1:n] − cj,3[1:n] ) (40k)
Ds,j
+ Sj,4 (cj,4[1:n] − cj,3[1:n] ) (40l)
dr Vj,4
Dcj,4 = Sj,4 (cj,3[1:n] − cj,4[1:n] ) (40m)
Ds,j
cj,{0:4},[1] = cj,init [1] (40n)
cj,{0:4},[n] = cj,init [1] (40o)
pmin [1] ≤ p ≤ pmax [1] (40p)
 
RT cj,max − cj,0 vmin [1] ≤ vbat ≤ vmax [1] (40q)
8j = v0bat + ln (40d)
F cj,0 imin [1] ≤ ibat ≤ imax [1] (40r)
 
N  k+1
[0] ≤ cj,{0:4} ≤ cj,max [1] (40s)
X Ak  2cj,0 2cj,0 k(cj,max −cs,j,0 )
+ −1 − c l + p ≤ τ [1]

F cj,max cj,max (2 c j,0 −1)1−k (40t)
k=0 j,max

where x SPM = p, pdc , ibat , vbat , 8p , 8n , ηp , ηn , cp,{0:4} ,


ibat
= ks,j cj,max c0.5 0.5 0.5
e (−cj,0[1:n] ) cj,0[1:n] (40e)
as,j As,j Ls,j cn,{0:4} , τ ∈ R18n+13 , 8p ∈ Rn+1 and 8n ∈ Rn+1 are the

VOLUME 7, 2019 178371

 D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

open-circuit-voltages of the cathode and anode respectively,

ηp ∈ Rn and ηn ∈ Rn are the overpotential voltages of the
cathode and anode respectively, and cp,{0:4} ∈ R5×(n+1) and
cn,{0:4} ∈ R5×(n+1) are the molar concentrations of active
material in the five discretized volumes of the cathode and
anode respectively.
Under this control design customer’s net load and optimal
control schedule for the BESS are shown in Fig. 16 (a) and (b)
respectively. The simulated battery current and voltage are
shown in Fig. 16 (c) and (e) respectively. The SPM based
controller expects to be able to reduce the peak load by
approximately 86 kW. The peak battery voltage reached
800 V and the dc current reached -125 A on discharge. Note
that the dc voltage is based on the anode and cathode voltage
functions which are in turn based on anode and cathode
concentration fractions. As these functions are different from
what the CRM uses, the results are not directly comparable.
As shown in Fig. 16 (f) and (g), the cathode concentration
fraction ranges from 0.50 at peak SoC to 0.87 at minimum
SoC while the anode concentration ranges from 0.76 at peak
SoC to 0.2 at minimum SoC. At low current densities there
is almost no difference between core and surface particle
concentrations. The control solution reduces the total elec-
trical bill from $52,080 ($50,000 demand, $2,080 energy) to
$47,754 ($45,682 demand, $2072 energy). The net effect is a
$4,325, or 8.31%, reduction in the electrical bill.
The SPM is structurally similar to the CRM, as illustrated
in the comparison of Fig. 9 and Fig. 15. However, the SPM
accounts for overpotential voltages in a different way from the
equivalent circuit models in the CRM. These differences are
not salient at the low sample rate in the example application
and hence the models appear to have very similar results.

IV. TEMPERATURE MODELS

Temperature is a critical factor to consider when controlling
BESS. Cell temperature can affect many of the parameters for
the SoC and SoH models discussed in Sections III and V. The
highest cell temperature can be the limiting factor for control
action in hot environments or under high power conditions.
Constraining temperature prevents over-temperature and, in a
few cases, under-temperature conditions which can shorten FIGURE 16. Results calculated with the SPM (a) net load with BESS power
control, (b) battery power, (c) battery SoC, (d) battery current, (e) battery
battery life or cause hazards such as thermal run-away. The voltage, (f) cathode concentration fraction, (g) anode concentration
following constraint enforces limits on temperature: fraction.

Tmin ≤ T ≤ Tmax (41) 

imax
 if T ≤ Tthr
where T is the battery temperature, Tmin is the minimum i0max = imax /2 if Tthr < T ≤ Tmax (42c)
battery temperature, and Tmax is the maximum battery
if T > Tmax

0

temperature.
Battery specification sheets will often define a lower max- where i0max is a dynamic charge current limit, Tthr is the tem-
imum charge rate at higher temperatures [92]. This require- perature threshold. This type of constraint is non-convex and
ment would specify a conditional dynamic charge limit based difficult to work with in optimal control design. Alternatively,
on if the battery temperature exceeds a given threshold, these restrictions can be implemented with affine constraints
an example of which is shown in: on current as in:
imin ≤ ibat ≤ imax (42a) m1 T + b1 ≤ ibat ≤ m2 T + b2 (43a)
ibat ≤ i0max (42b) m3 T + b3 ≤ ibat ≤ m4 T + b4 (43b)

178372 VOLUME 7, 2019

D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

TABLE 8. Summary of temperature models.

sent. Section IV-A explains the physical mechanisms under-

lying heat generation and transfer for batteries. Section IV-C
then introduces several specific modeling approaches that
can incorporate temperature into optimal control decisions.
Table 8 shows a summary of the benefits and tradeoffs of
the different temperature models as they apply to optimal
controller design.

A. HEAT GENERATION, CONSUMPTION, AND TRANSFER

The temperature of the battery is a function of the rate of
heat generated by the battery during operation (Jin ) and the
rate of heat lost to the environment (Jout ). Heat is gener-
ated or consumed by an electrochemical cell in three ways:
FIGURE 17. Notional examples of temperature dependent current limits. change in entropy, overpotential losses, and resistive heat-
ing. Changes in entropy from the electrochemical reactions
reversibly generate and consume heat within cells. This pro-
where m1−4 and b1−4 are the slopes and intercepts of the tem- cess is referred to as reversible heat generation because the
perature dependent current constraints. Fig. 17 shows how the heat generated during charge or discharge is consumed during
constraints in (42) and (43) enclose different feasible regions. the reverse reaction. Charging a battery can be endothermic
This kind of limit can be imposed on dc power instead of (e.g., some types of lithium batteries in specific ranges of
current [93]. SoC), or exothermic (e.g., lead-acid batteries) [10], [65], [88],
Battery temperature models are based on how much heat is [100], [102], [103]. When losses are considered, we reintro-
generated in the cell, and how much heat is lost to the envi- duce the equivalent circuit models outlined in Section III-B.
ronment. As controllers must balance accuracy with model Overpotential losses result from the kinetic and mass trans-
complexity, we cannot use the high order finite-element- port aspects of the chemical recreation which are modeled
models used in simulation based design, like in [94]. Instead by the resistor-capacitor ladder in the 2nd order equivalent
controller models choose a few critical temperatures to repre- circuit. The voltage drop across these elements are v1 and

VOLUME 7, 2019 178373

 D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

or discharging. However, the accuracy improvement may not

warrant the increased computational complexity.
Heat can be transferred between the battery and the envi-
ronment through conduction, radiation, and convection. Heat
conduction is proportional to the temperature difference
while heat radiation is proportional to the temperature of
the surface raised to the 4th power according to Stefan-
Boltzmann’s law [106]. The combined heat loss function is
shown in:
Jout = U (Tenv − T ) + σem (Tenv
4
− T 4) (45)
FIGURE 18. Computed full cell 1S from individual electrode 1S for three
types of lithium-ion batteries, replotted with data from [104]. where Jout is the rate of heat loss (W ), U is the battery’s
thermal transmittance with its environment (W/K ), Tenv is the
environmental temperature (K ),  is the Stefan-Boltzmann
v2 respectively. Alternatively, if the SPM or P2D model is constant (5.6 × 10−8 Wm−2 K−4 ), σem is the emission ratio
used for a SoC model, the overpotential voltages ηp and with respect to the ideal (0.95 is common for plastics in a
ηn can be used in place of v1 and v2 . Resistive or joule variety of battery designs [106]).
heating losses result from the power dissipated to the battery’s In an unregulated environment Tenv can be forecasted
internal resistance. Combining these three sources of internal based on local weather data. In a temperature controlled
heat generation yields : environment, it can sometimes be assumed that Tenv is a
constant. The heat transfer from radiation is normally much
1S(ς)
Jin = ibat T + (v1 + v2 )ibat + R0 (ibat )2 (44) smaller than then heat conduction meaning that it can be
nmol F ignored in many systems. A simplifying assumption is that
where Jin is the rate of heat generation (W ), ibat is the bat- the airflow rate is constant, thereby yielding a constant U .
tery current (A), T is battery temperature (K ), 1S is the However, in some cases variable speed fans can be integrated
change in entropy (1S = nmol F(∂voc /∂T )), nmol is the into the optimal control design. Under variable airflow condi-
number of electrons per reaction, F is the Faraday constant tions the rate of heat transfer is described by Nusselt number
(-1/96,485 Coulombs per electron), v1 and v2 are the dynamic (Nu) which itself is a function of Reynolds number (Re)
battery voltages from the equivalent circuit (V), R0 is the and Prandtl number (Pr). One example of this relationship,
battery internal resistance (), and voc is the open-circuit- from [107], is shown in:
voltage (V). The total change in entropy in a battery can ρ u∞ lmm
change drastically as a function of SoC which can be difficult Re = (46a)
µ
to model for the purposes of control design. The change in
Nu = Cl Reθ Pr 3
1
entropy over the domain of SoC was calculated from precise (46b)
measurements of ∂voc /∂T for a selection of lithium-ion bat- Nu ktc
U = (46c)
tery types as shown in Fig. 18 [104]. From these data we can lmm
contrast the low entropic heat generated on discharge from where ρ is the fluid’s density, u∞ is the unobstructed velocity
lithium-iron-phosphate (LFP) batteries to the relatively high of the fluid, lmm is the characteristic length, µ is the dynamic
entropic heat from lithium-cobalt-oxide (LCO), especially in viscosity, Cl and θ are empirically derived model parameters,
the range of 10% to 40% SoC. Lithium-manganese-oxide and ktc is the thermal conductivity of the fluid.
(LMO) batteries in further contrast change from generating Table 9 shows the thermodynamic constants associated
heat (negative 1S) to consuming heat (positive 1S) when with air and water under standard temperature and pressure.
passing 50% SoC on discharge. Depending on the battery As water has a much higher density and thermal conductivity,
chemistry, and the range of operational SoC, the 1S(ς ) some BESS designs include water cooling systems [108].
function may be neglected entirely, or approximated by a For a cylinder in cross-flow, the characteristic length lmm
constant, a linear function, a quadratic function, or cubic equals the diameter (18.63 mm for an 18650-type cell). The
function, or even a cubic spline [56], [105]. Accurate yet parameters Cl and θ, shown in Table 10, are properties of
simple models for changes in entropy that controllers can the geometry of the fluid flow over the battery surface and
use to predict temperature are an underdeveloped area that change with the Reynolds number. Together, these material
warrants additional research. properties yield the functional relationship between air speed
Most of the heat generated in a cell, especially in high thermal transmittance shown in Fig. 19.
power applications, comes from the resistive heating term. The complex relationships described in (46) impact control
Because of this, some choose to ignore overpotential losses design in several ways. First, heat transfer rate increases with
and the thermochemistry entirely [95]. Just as with the equiv- increasing fluid velocity u∞ . However, there are diminish-
alent circuit, the heat generation can be calculated with differ- ing returns meaning that the marginal improvement in heat
ent sets of parameters depending on if the battery is charging transfer decreases with increased fluid velocity. Hence there

178374 VOLUME 7, 2019

D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

TABLE 9. Thermodynamic properties of common battery cooling air conditioner is measured in the energy efficiency ratio
fluids* [107].
(EER), which is the ratio of Btu per hour cooling to power
input (W). The EER ranges from roughly 10-20 for high
efficiency units [109] which, given that 1 Btu per hour =
0.293 watts, correlates to a energy efficiency (η HVAC ) of
300-700%. A modified version of the heat dissipation equa-
tion for a BESS enclosure is shown in:
TABLE 10. Fluid flow geometry constants for a cylinder in
cross-flow [107].
JEN,out = U EN (Tenv − T EN ) + σem (Tenv
4
− T 4EN )
+ σem p irr − η HVAC p HVAC (47)

where U EN is the thermal transmittance between the enclo-

sure and the environment, T EN is the enclosure tempera-
ture, p irr is solar irradiance, p HVAC is the ac power load
of the HVAC unit, and η HVAC is the HVAC’s energy effi-
ciency. This approach assumes constant airflow and temper-
ature in the HVAC’s heat exchanger. By modeling the heat
exchanger temperature and fan, we can improve the con-
troller’s accuracy predicting temperature management costs.
The expanded enclosure thermal model is shown in:

J EN,out = U EN (Tenv − T EN ) + σem (Tenv

4
− T 4EN )
+ σem p irr +U EX (u∞ ) (T EX −T EN ) (48a)
∂T EX
C EX = U EX (u∞ ) (T EN − T EX ) − η HVAC p HVAC
∂t
(48b)
u∞ = η fan p fan (48c)
FIGURE 19. Thermal transmittance for a 18650 battery cell in cross flow
as a function of air speed [107]. where U EX (u∞ ) is the thermal transmittance between the
HVAC heat exchanger and the air, which is a function of the
airflow u∞ , T EX is the heat exchanger temperature, C EX is
is likely to be an optimal, non-zero flow rate that effectively the heat exchanger’s heat capacity, p fan is the fan power, and
transfers heat while not consuming too much power to move η fan is the fan’s efficiency (m s−1 kW−1 ).
air. A fan controller can be implemented to optimize battery Many HVAC systems are controlled using thermostats,
temperature along with charge/discharge [101]. which activate heating or cooling modes when outside a set
temperature range. The simplest thermostat implementation
B. ENCLOSURE THERMAL MODEL is shown in:
The simplest enclosure thermal model is implicit in the 
assumption of constant environmental temperature. This is p cool T EN > T high

valid if the BESS is small and installed in a temperature- p HVAC = 0 T low ≤ T EN ≤ T high (49a)
controlled space. Temperature forecasts can be used in envi- p heat T EN < T low


ronments where temperature is weather dependent. As there
is generally thermal separation between the weather depen- where p cool is the power of the HVAC when in cooling mode,
dent environment and the environment that the batteries are p heat is the heating power of the HVAC when in heating
operating in, we can model these temperatures separately. mode, and T high and T low are the high and low environment
For a given BESS and environment, the heat transfer temperature limits respectively. To limit the on/off cycling
rate between the enclosure and the environment U EN can frequency, the mode will often stay latched for a set dura-
be empirically calculated with heating ventilation and air tion, or until the desired temperature is reached. However,
conditioning (HVAC) systems off, and over their range of this operational mode is recursive, meaning it is difficult
control. We may also consider the effect of solar heating to incorporate into a computationally efficient optimal con-
which is proportional to irradiance. The HVAC power draw troller design. If it is feasible in the design of the HVAC
can also be calculated under these conditions, and the result- system, HVAC power (p HVAC ) and/or fan power (p fan ) can be
ing functions can be included in the BESS model. Further, decision variables available to the controller. This modeling
an HVAC system can transfer a greater amount of heat approach enables optimal HVAC control scheduling, includ-
from the enclosure to the environment, or vice versa, than ing pre-cooling batteries [101] or pre-heating batteries [98]
it requires in electrical energy. The efficiency of a room to prepare for usage later in the control horizon.

VOLUME 7, 2019 178375

 D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

FIGURE 20. Cell thermal model w/ a lumped-volume, resistive-heating,

and conduction-cooling.
FIGURE 21. Enclosure model with lumped-air-volume and HVAC
efficiency cooling.

C. TEMPERATURE MODEL TYPES

In this section we develop several increasingly complex, accuracy as shown in:
battery-cell thermal models to illustrate the different options
∂T 0
for thermal model design. To build these models we pull CT0 = (v1 + v2 )ibat + R0 (ibat )2 (52a)
together the physical mechanisms discussed in Sections IV-A ∂t
and IV-B into systems of constraints. The simplest, and most + U 0 (T − T 0 ) (52b)
widely used, model is to only represent a single temperature ∂T
CT = U 0 (T 0 − T ) + U (Tenv − T )
(often the hottest cell), considering only resistive heating and ∂t
conduction-based cooling. An example of this type of model + σem (Tenv4
− T 4) (52c)
is shown in Fig. 20 and in:
where CT0 is the heat capacity of the internal mass of the cell,
∂T T 0 is the internal temperature, and U 0 is the thermal transmit-
CT = R0 (ibat )2 + U (Tenv − T ) (50)
∂t tance between the internal mass to the surface. As the model
now distinguishes between surface and internal temperature,
where CT is the heat capacity of the lumped-volume.
CT is now the heat capacity of the surface of the cell, T is
In this lumped-volume model the measurable surface tem-
the surface temperature, and U is the thermal transmittance
perature is assumed to be the temperature throughout the cell
between the battery surface and its environment. The result-
[95], [96]. Note that while we have depicted the cell geom-
ing model structure is illustrated in Fig. 22.
etry as cylindrical, this approach works equally well for
pouch or prismatic cells. Note that when only the hottest
cell is represented, imposing a low temperature constraint
is unnecessary and potentially misleading. Instead, the low
temperature limit is enforced either by battery selection at the
design stage (i.e. picking a battery chemistry that is suitable
for its environment) or by designing a HVAC system with a
thermostat that regulates environmental temperature.
Where the BESS includes an enclosure with a controllable
HVAC system, the model can include an additional state vari- FIGURE 22. Cell thermal model w/ internal and surface-volumes, resistive
and overpotential heating, and conduction/radiation cooling.
able for the enclosure temperature [98] as in Fig. 21 and in:

∂T Building on this framework we can add additional internal

CT = R0 (ibat )2 + U (T EN − T ) volumes, entropy based heating, and convection cooling as a
∂t
∂T EN function of air velocity as shown in:
C EN = Ncell (U (T − T EN )) + U EN (Tenv − T EN )
∂t ∂T10
− η HVAC p HVAC (51a) CT0 1
∂t
= U10 (T20 − T10 )
where C EN is the heat capacity of the BESS enclosure, U EN
1S(ς)
 
is the thermal transmittance between the enclosure and the + V1 ibat T10 + (v1 + v2 )ibat + R0 (ibat )2
environment, and Ncell is the number of cells in the enclosure. nmol F
This model assumes constant or no airflow. 0
∂T 0
[2:K ]
While temperature measurement is performed on the sur- CT [2:K ]
∂t
face of batteries it is a better practice to constrain opera- 0 0 0
= U[2:K ] (T[3:K +1]) − T[2:K ] ) (53a)
tion based on limiting the maximum internal temperature
1S(ς)
 
0 2
[51], [99]. The internal temperature can be estimated based + V[2:K ] ibat T[2:K ] + (v1 + v )i
2 bat + R (i
0 bat )
on the surface temperature and the battery’s operation. nmol F
0 0 0
We can also include battery over-potential heating to improve + U[1:K −1] (T[1:K −1] − T[2:K ] ) (53b)

178376 VOLUME 7, 2019

D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

∂T
CT
∂t
= U (u∞ ) (Tenv − T ) + UK0 (TK0 − T )
+ σem (Tenv
4
− T 4) (53c)
where TT [1:K ]0 are the K internal temperatures, CT [1:K ]0 are
0
each internal volume’s heat capacity, V[1:K ] are the volumetric
fractions of each internal volume normalized to the total
internal volume, and U[1:K0 FIGURE 24. Enclosure model with, lumped-air-volume, heat-exchanger
] are the thermal transmittances temperature, HVAC efficiency and fan speed based cooling.
between internal volumes. Note that for notation simplicity
the surface temperature TK0 +1 = T in (53b). The resulting
model structure is illustrated in Fig. 23. TABLE 11. Example battery system temperature model parameters.

FIGURE 23. Cell model with multiple-internal-volumes,

resistive/overpotential/entropy-heating, and
conduction/radiation/convection-cooling.

To take advantage of the convection cooling term,

the enclosure model can be further developed to include fan
power, air velocity, and the temperature of the HVAC heat-
exchanger. To accomplish this, we replace Tenv with T EN in
(53c) and add the additional constraints shown in:
1) TEMPERATURE MODEL APPLICATION
∂T EN  
For this application we solve the optimal control problem in
C EN = Ncell U (u∞ ) (T −T EN )+σem (T 4 −T 4EN )
∂t Section II using a thermal model. However, in this section
+ U EN (Tenv − T EN ) + σ EN (Tenv
4
− T 4EN ) we assess how the control changes if it is in a very hot envi-
+ U EX (u∞ ) (T EX − T EN ) (54a) ronment. In some regions, the temperature can commonly
∂T EX reach 43.3 ◦ C (110 ◦ F) during the day. BESS in such an
C EX = U EX (u∞ ) (T EN − T EX ) − η HVAC p HVAC environment are generally installed in enclosures with HVAC
∂t
(54b) systems. Given this environment, we determine an optimal
control schedule for both the BESS power and the HVAC
u∞ = η fan pfan (54c)
system power using the parameters in Table 11. The modified
where T EX is the heat-exchanger temperature, C EX is the objective and constraints, in addition to those for the CRM
heat exchanger heat capacity, U EX (u∞ ) is the air velocity defined in Section III-B.1, are shown in:
dependent thermal transmittance between the air and heat-
exchanger, p fan is the fan power, and η fan is the fan efficiency. min 1t w> (l + p + p HVAC ) + ντ (55a)
The resulting model structure is illustrated in Fig. 24. x T ∈R9n+5
There are many useful combinations of these models. For subject to:
example, a controller may want to have a more detailed cell ..
model and a less detailed enclosure model or vice-versa. . in addition to the constraints in (28)
Alternatively, these models can be customized to a specific l + p + p HVAC ≤ τ (55b)
cell design or enclosure architecture. The ‘‘Big cell’’ mod- CT DT = R0 (ibat )2 + U (T EN[1:n] − T[1:n] ) (55c)
eling assumption is commonly used but the ‘‘Short-board
C EN DT EN = K H (T[1:n] − T EN[1:n] )
effect’’ and ‘‘One-by-one calculation’’ can be used in ther-
mal modeling as well. A similar model extension to the + U EN (Tenv − T EN[1:n] )
enclosure would be to represent a finite number of internal −η HVAC p HVAC (55d)
volumes. Each cell would reside within a volume and the T[1] = T0 (55e)
heat transfer would only depend on that volume’s temper-
T EN[1] = T0 (55f)
ature. Each of these options greatly increases model com-
plexity with unknown, perhaps limited, benefits to controller T ≤ Tmax [1] (55g)
performance. [0] ≤ p HVAC ≤ p HVAC-max [1] (55h)

VOLUME 7, 2019 178377

 D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

where x T = {p, pdc , ibat , vbat , voc , ς , T, p HVAC , T EN , τ, } ∈

R9n+5 , T ∈ Rn+1 is the temperature of the hottest cell at each
time step, T EN ∈ Rn+1 is the enclosure temperature, and
ibat ∈ Rn is the dc current. The environmental temperature is
assumed to be sinusoidal, with a period of 24 hours, a peak
of 45◦ C at 3:00 pm, and a magnitude of 2.5◦ C. The formal
expression for the temperature is shown in:
2π 1t
 
Tenv = 2.5 cos k − 15 +42.5 ∀ k ∈ {1, 2, . . . n}
24
(56)
where Tenv ∈ Rn is the environmental temperature at each
time step.
The net load achieved over the control horizon from the
combined SoC-Thermal model is shown in Fig. 25 (a). The
optimal control schedule calculated is shown in Fig. 25 (b).
The HVAC power schedule is shown in Fig. 25 (d) and the
environmental, battery, and enclosure temperature trajecto-
ries are shown in Fig. 25 (e). The controller can anticipate
a period of high temperature and pre-cool the enclosure, and
hence the battery, to achieve the desired schedule. Note also
that the pre-cooling takes place during the off-peak electricity
pricing period. The magnitude and duration of the HVAC
cooling is precisely tuned such that the battery’s temperature
reaches its limit (45 ◦ C) exactly at the end of the sched-
ule. Note also that the power profile no-longer preferentially FIGURE 25. Results from control incorporating temperature model:
(a) net load with BESS power control, (b) battery power, (c) battery SoC,
charges during off-peak times. This is a result of the quadratic (d) enclosure ac power, and (e) battery, enclosure, and environmental
increase in temperature from high rate charging that generates temperatures.
too much heat for the system to transfer to the environment
cost-effectively.
The control solution reduces the total electrical bill introduce and assess empirical stress factor based models for
from $52,080 ($50,000 demand, $2,080 energy) to $48,001 accurate degradation modeling. Linearizing and simplifying
($45,871 demand, $2,130 energy). Within the energy bill, the detailed empirical degradation model allows us to calcu-
the energy required to cool the battery accounts for $51. The late several norm-based regularization factors that efficiently
net effect is a $4,079 (7.83%) reduction from the baseline incorporate degradation into optimal control objectives. Last,
electrical bill, or a $53 (0.11%) increase in the electrical we cover physical degradation models based on intercala-
bill calculated using only the CRM. The more important tion stresses and two different side-reactions in lithium-ion
comparison is that if we model battery temperature in this batteries.
environment under the control solution developed using only We use the terminology beginning-of-life (BoL) to denote
the CRM, the hottest battery reaches a peak temperature the conditions when the battery is new, end-of-life (EoL) to
of 55.6 ◦ C. By incorporating a thermal model into the con- denote the conditions when the battery can no-longer reliably
troller, we can plan control actions to maintain defined tem- supply energy services, and state-of-life (SoL) to denote the
perature limits. conditions, between BoL and EoL, that the battery is in at
a given state. The EoL conditions are often specified by the
V. DEGRADATION MODELS battery or BESS manufacturer as a part of a warranty. Because
As batteries age with time and use, their energy storage and of this lack of standardization, SoH can be defined in many
supply capabilities degrade until they no-longer meet the ways (e.g., based on changes in capacity [114], resistance,
requirements of their designed services. When degradation round trip efficiency, etc.). We use a more general definition
is included in optimal control, it tends to rely on empirical of SoH (%), represented in:
degradation models that abstract many of the physical pro- y BoL − y SoL
cesses in favor of model simplicity. However, there are several % =1− (57)
studies that use the SPM or a simplified P2D to incorpo- y BoL − y EoL
rate physical degradation models into a controller design where % is the SoH of the battery and y is a critical param-
[46], [81], [110]–[113]. eter for the battery to reliably supply services. In (57),
This section first establishes definitions for state-of-health SoH is defined as the ratio of a specific parameter’s (or
(SoH) and how they fit into optimal control. We then combination of parameters) movement from its initial state

178378 VOLUME 7, 2019

D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

at BoL to its final state at EoL. In this context, SoH can where i is the interest rate, n is the number of compounding
represent movement in energy capacity (kWh), available periods between SoL and EoL, and tcomp is the duration of
energy (kWh), charge capacity (Ah), available active material each compounding period. This is meant to illustrate only
(e.g. through ‘‘loss-of-lithium’’ in mols), coulombic effi- one of many methods available for discounting future cost.
ciency (%), or internal resistance (). Available energy is a It is sensible that control design using this method reflects
combination of available charge/discharge power, and energy whatever financial structure and assumptions are used for
capacity that is defined very precisely in the electric vertical projected EoL costs.
context [114] but analogs can be imagined for energy storage Alternatively, a controller can be designed to maximize
applications as well. Using this definition, no matter what value while enforcing a designed or warranted service life.
parameter is used, and whatever the BoL and EoL conditions This method does not include an additional cost term in the
are specified, % at BoL always equals 1, and % at EoL always objective and instead includes an additional constraint on the
equals 0. As we are focused on controller design, this def- average rate of degradation, as shown in:
inition does not account for ‘‘rejuvenation’’ cycles wherein
%
lead-acid and some types of flow batteries can recover some %̇ ≥ − (60)
loss of SoH. Lwar − L
For the purposes of control design, we can assume that
where Lwar is the total warranty life (e.g., 15 years), and L
the change in model parameters from degradation over any
is the current life (years that the BESS has been in service).
forward-looking control horizon is extremely small. That is,
Critically, the degradation rate should be allowed to temporar-
absolute changes in parameters from degradation happen over
ily exceed the rate at which the BESS would reach EoL before
the course of months or even years, while controllers operate
the warranted service life as this allows for periods of rest to
over hours or days. For this reason, from the perspective of
counterbalance period of high utilization. If this method is
control design, it is unimportant which parameter is used
used, it is important to account for how the controller should
to calculate SoH. The rate of degradation, in contrast, can
transition operation past EoL as (60) is infeasible if L ≥ Lwar .
change quickly and is a critical factor in determining optimal
Note that (60) can be imposed as a soft constraint, with a
control. Hence, rather than modeling SoH, we model the rate
slack variable subtracted from the limit and maximized in
of degradation directly as a calculated variable.
the objective. This approach can handle infeasibility at the
There are at least two ways to incorporate the rate of degra-
expense of additional decision variables, which can be helpful
dation into optimal control design. The first way is to add
when more complex degradation models are used.
incremental battery replacement/refurbishment cost in the
A useful reformulation of this is for a manufacturer to
objective [30]. The second method is to constrain operation to
supply a ‘‘warranty life curve’’ as shown in Fig. 26. This
a maximum degradation rate to ensure a warranty period [28].
curve has a maximum warranty life and a function that
The following is a detailed introduction to these two methods.
describes how the warranty period would be shortened based
When batteries reach EoL, they can be replaced with
on BESS operation increasing a supplied degradation metric
new batteries that restore the system’s functionality to BoL
(e.g., cycles as in [115]). This curve may or may not be
conditions. In certain cases, the old batteries can be resold/
accompanied by an equation to calculate the degradation
re-purposed in a new application. The net costs predicted to
metric as it is often described by just a few points to prevent
be incurred at EoL, denoted by C EoL , provide a quantitative
reverse engineering. The warranty life can be interpreted in
estimate of how much the controller should weight battery
the context of control as the reciprocal of the rate of degrada-
degradation. The cost incurred through battery degradation is
tion, as in:
calculated in:
∂% 1 %
fb = C EoL = C EoL %̇ (58) %̇ = − ≥− (61)
∂t DLwar Lwar − L
where fb is the cost of the battery degradation over a full
control horizon, C EoL is the net cost at EoL, % is the present where D : R 7→ R is the warranty life curve supplied by
SoH, and %̇ is the average degradation rate over the control the battery or BESS manufacturer. This formulation allows a
horizon. generic warranty life curve to be implemented as a constraint
As the cost incurred in (58) is the present value of a pre- into a BESS controller.
dicted future cost, it is possible to apply a discount rate based This section outlines various models for calculating the
on an assumed interest rate. The number of compounding average rate of degradation for use in optimization. We adapt
periods would then be estimated linearly from the current a stress factor model used for life prediction of lithium-ion
SoH, the average rate of degradation, and an assumed com- cells for use in control design. We then illustrate how, through
pound period, as shown in: a series of operational assumptions, this stress factor model
can be reduced to simple norm-based regularization. Last,
fb = (1 + i)−n C EoL %̇ (59a) we introduce several physical degradation models. Table 12
%
 
1 shows a summary of the types of degradation models dis-
n= (59b)
tcomp %̇ cussed in the following sections.
VOLUME 7, 2019 178379
 D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

TABLE 12. Summary of Degradation models.

being the superposition of the two [132]. Calendar aging

models are functions of time, average SoC, and average
temperature and impact SoH whether or not the battery is
charged or discharged. Cycle aging models are based on cycle
SoC, current, cycle depth-of-discharge (DoD), and cycle tem-
perature. Models based on current (or C-rate), such as the
models presented in [133] and [134], generally work best
for constant current cycling performed in laboratory experi-
ments and have unknown accuracy in application that require
variable charge or discharge rates.
It is common to represent degradation based on an
exponential decay function of calendar and cycle degrada-
FIGURE 26. Notional example of a manufacturer supplied degradation tion [121], as shown in:
curve (curve value = D).
% = e−fd (62)
A. EMPIRICAL DEGRADATION MODELS
Empirical battery degradation models can be classified as where fd is the aggregate degradation stress factor based on
either calendar aging or cycle aging, with total degradation a combination of calendar life stress factors, and cycle life

178380 VOLUME 7, 2019

D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

stress factors. We can model the value of the aggregate degra- circumstances, the optimal solution, assuming only one cycle,
dation stress factor as an additional variable in our system yields two or more cycles. Similarly, the optimal solution
representation. assuming two cycles can often yield an optimal schedule that
A rainflow cycle counting algorithm originally developed includes just one cycle, or cycles with different boundaries
for material degradation [135] is widely used for accurate than expected. The second approach is to discretize the con-
cycle-life modeling [52], [122], [123]: trol schedule as demonstrated in [115], [124]. By breaking the
N available range of SoC and current into a number of discrete
X
fd = St Sς ST + wi Sδ Sς ST (63) states it allows the controller to map each state transition
i=1 onto a piecewise linearized degradation curve. This approach
has the benefit of accuracy of the degradation function at
where N is the number of cycles in the control horizon, i is a
the cost of precision of the control solution and computation
cycle index variable, wi is a binary variable indicating a full
time. The third workaround is to linearize the degradation
cycle or a partial cycle, and each stress factor is shown in:
rate around assumed static operational conditions, including
St = kt t (64) cycles. Doing this, it can be found that the rate of degra-
Sς = ekς (ς−ςref ) (65) dation can be written in the form of a regularization term.
Tref This third approach is discussed in the following section.
ST = ekT (T −Tref ) T (66) Lastly, an apparent solution is presented by Shi et al, who
Sδ = a δ 4 + b δ 3 + c δ 2 + d δ + e (67) first prove the convexity of the rainflow counting algorithm
and then demonstrate a subgradient algorithm for efficient
where t is time, St is the time stress factor, ς is SoC, Sς is
optimal control [125]. This method works by recognizing
the SoC stress factor, T is temperature, ST is the temperature
that every charge (and discharge) action belongs to either one
stress factor, δ is DoD, and Sδ is the DoD stress factor. The
charge half cycle or two charge half cycles if it is at the time
parameters Tref , ςref , kt , kς , kT , a, b, c, d, and e enable
boundary between two cycles. The cost of a charge action at
their associated stress factors to be tuned to specific bat-
the boundary can be mapped from the cycle depth of either of
teries. Degradation models that do not use rainflow cycle
its member half cycles. Hence the subgradient algorithm can
counting often make duty-cycle profile assumptions such
avoid the calculation of the number of cycles entirely, instead
as in [136], [137].
adding the cost associated with the member half cycle to the
Here we could extend the short board and cell-by-cell
subgradient of the charge action.
modeling approaches introduced in Section III-B to the dis-
tribution of degradation rates within a battery string or pack.
1) DEGRADATION AS REGULARIZATION
However, from a control perspective, representing the max-
imum and minimum SoH in a string is less critical than In machine learning, regularization is commonly used to
for either SoC or Temperature because particularly low SoH prevent a model from overfitting data. Here we use similar
cells can be replaced during regular maintenance, and hence methods to prevent our controller from over-using batteries.
would not limit operation. For this reason, the ‘‘big cell’’ In this section we derive several different kinds of regular-
representation of string level degradation is generally the ization terms based on the stress factors described above.
most appropriate for optimal control applications. While most of the degradation stress factors are nonlinear
As the controller objective is to minimize the change in functions, their first-order Taylor series approximations can
SoH, we can take the derivative of (62) to obtain: be reformulated as the norms of specific decision variables.
The simplest approach to calculating the rate of degra-
∂fd −fd dation (%̇) is to linearize it to an assumed cycle depth-
%̇ = − e = −kt Sς ST e−fd (68)
∂t of-discharge, temperature, and average SoC. Under these
yielding the form of SoH used in a controller model. assumptions, the degradation rate can be written:
Modeling SoH in this way presents a fundamental chal-
|pe |
lenge. The rainflow counting algorithm in (63) is recursive %̇ = 1
(69)
in that, under most conditions, we cannot determine the (1 + ηe )Lcyc Qcap
number or time of each cycle within an schedule. When per- |ibat |
%̇ = (70)
forming a rainflow counting algorithm on a known schedule, (1 + 1
ηc )Lcyc Ccap
the schedule is broken into many smaller pieces that add
up to the total degradation. However, this schedule split- where p is BESS ac real power, ibat is the battery current, ηc is
ting cannot be done a priori and hence is very difficult the coulombic efficiency, Lcyc is the rated cycle-life to EoL,
for optimization algorithms to work with. We discuss three Qcap is the energy capacity, and Ccap is the charge capacity.
imperfect workarounds and one apparent solution to this Under these narrow conditions, degradation is proportional
problem. First, there are some cases where the time windows to the absolute value of the battery power as shown in (69)
for each cycle are predetermined (e.g., daily cycling). This when using the ERM [28], [30], [116], [138] or to the absolute
makes the rainflow counting algorithm trivially simple and value of the battery current as shown in (70) when using
easy to implement in optimization. However, under some the CRM [46], [74]. A modification to this approach is to

VOLUME 7, 2019 178381

 D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

establish a power or current threshold above which the linear in [117]–[119] to minimize heat generation in hybrid vehicle
cost increases as in [33]. Another modification is to multiply energy management optimization.
the absolute value of power or current by a stress factor based Further, assuming a constant battery voltage (vbat = v0 ),
on temperature and charge/discharge rate as in [51], [116]. (77) can be reformulated using ac power instead of dc current.
Note that, when adding (69) or (70) up over a discrete control Assuming the ac/dc conversion model in (24), with φb = 0,
horizon, this form of degradation is equivalent to applying a the minimum heat generation regularization is shown in:
`1 norm power regularization, or `1 norm current regulariza- fb (p) = 5T ||p||22 (78a)
tion, as shown in:
1t C EoL KT R0 φm
2 2
5T = (78b)
fb (p) = 5 cyc ||p||1 (71a) CT v0
1t C EoL where the regularization weight 5T in this equation has units
5 cyc = (71b)
(1 + η1e )Lcyc Qcap of $/kW 2 .
At an assumed static average SoC (ςavg ), the the derivative
fb (ibat ) = 5 cyc ||ibat ||1 (72a)
of the degradation rate with respect to average SoC is con-
1t C EoL
5 cyc = (72b) stant, as shown in:
(1 + η1c )Lcyc Ccap ||ς||1
ςavg = (79)
The regularization weight 5 cyc has units of $/kW or $/A n
depending on which equation it is in because of the units of T =T0
the relevant decision variable. d 2% ∂  ˙ −fd  t=t0
Kς = = −fd e (80)
At an assumed static temperature, the derivative of degra- dςavg dt ∂ςavg δ=δ0
dation rate with respect to temperature is constant (KT ) as ς =ς0
shown in (73). Ignoring the reversible heat generation, over- where Kς is the partial derivative of degradation rate with
potential heating, and assuming that battery temperature and respect to SoC, ςavg is the average SoC, and n is the number
environmental temperature are very close, the derivative of of steps in the discrete control horizon. Multiplying both
temperature in (50) reduces to simply the resistive heating sides by ∂ςavg and, assuming piecewise constant values for
term, as in: SoC, integrating yields an approximation for the average
degradation rate:
T =T0
∂ 2% ∂  ˙ −fd  t=t0 %̇ ≈
Kς
||ς||1 (81)
KT = = −fd e (73)
∂T ∂t ∂T δ=δ0 n
ς=ς0 This form of degradation is equivalent to applying a
∂T R0 2 `1 norm SoC regularization to the objective function as
= i (74) shown in:
∂t CT bat
where KT is the partial derivative of degradation rate with fb (ς ) = 5ς ||ς||1 (82a)
respect to temperature, T is the battery temperature, t is time, 1t C EoL Kς
5ς = (82b)
R0 is the battery ohmic resistance, and CT is the battery’s n
total heat capacity. From (73) and (74), we obtain the second where the regularization weight 5ς in this equation has units
derivative of degradation: of $/(%SoC).
KT R0 2 At an assumed static cycle DoD, the derivative of degrada-
%̈ = i (75) tion rate with respect to DoD is constant, as shown in:
CT bat
δ = max(ς) − min(ς) = ||ς ||∞ +||1 − ς ||∞ −1 (83)
Assuming piecewise constant values for current, integrat-
ing (75) yields an approximation of the average degradation T =T0
rate: ∂ 2% ∂  ˙ −fd  t=t0
K DoD = = −fd e (84)
1tKT R0 ∂δ ∂t ∂δ δ=δ0
%̇ ≈ ||ibat ||22 (76) ς=ς0
CT
where K DoD is the partial derivative of degradation rate with
Again note that, when added up over the control horizon,
respect to DoD, and δ is the DoD. Multiplying both sides by
this form of degradation is equivalent to applying a `2 norm-
∂δ and, assuming piecewise constant values for SoC, integrat-
squared current regularization to the objective function as
ing yields an approximation for the average degradation rate:
shown in:
fb (ibat ) = 5T ||ibat ||22 (77a) %̇ ≈ K DoD δ = K DoD (||ς||∞ + ||1 − ς||∞ − 1) (85)
1t C EoL KT R0
2 This form of degradation is equivalent to applying an `∞
5T = (77b) norm ς and 1 − ς regularization to the objective function as
CT
shown in:
where the regularization weight 5T in this equation has
fb (ς) = 5 DoD ||ς ||∞ + ||1 − ς||∞


units of $/A2 . This form of degradation cost has been used (86a)

178382 VOLUME 7, 2019

D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

δSEI
5 DoD = C EoL K DoD (86b) % SEI = 1 − (89b)
δSEI,EoL
Note that the −1 can be omitted from ς DoD in this formula-
tion because, as a constant, it would not affect the minimizers where MSEI is the molar volume of SEI reaction products,and
of the optimization. This degradation cost has been applied ρSEI is the density of the SEI layer. It should be noted that
to BESS in a daily energy market arbitrage application [120]. there are many other ways of modeling the growth of the SEI
The regularization weight 5 DoD in this equation has units of layer [129].
$/(%DoD). If capacity is the critical parameter for operation, then we
Now that the partial derivatives have each been derived, use the loss-of-lithium to calculate the rate of change in SoH
we can combine them to yield a function for total degrada- as shown in:
tion. As the current based cycle-counting and heat-generation
∂LSEI
degradation functions require fewer assumptions than their = JSEI As,n (90a)
ac power based counterparts, we use (72a) and (77) instead ∂t
LSEI
of (71) and (78) though either option produces a viable esti- % SEI = 1 − (90b)
mate of total degradation. The formulation for linearized total LSEI,EoL
degradation cost is shown in: where LSEI is the lost lithium content, and As,n area of the
fb (ibat , ς ) = 5 cyc ||ibat ||1 + 5T ||ibat ||22+ 5ς ||ς||1 negative electrode.
Another side-reaction to consider is lithium-plating, which
+ 5 DoD ||ς ||∞ + ||1 − ς||∞


(87)
can occur under adverse charging conditions or as a result
of accidental overcharge [81], [111]. In this case, the rate of
B. PHYSICAL DEGRADATION MODELS
change in SoH can be calculated as the magnitude of the side-
Physical degradation models have already been reviewed
reaction over-potential if it is negative as shown in:
in [85], [139], [140]. These models are built on top
of the concentration-based SoC model type discussed in ηsr = φ1,n − φ2,n − 8sr − F Jsr Rfilm (91)
Section III-C. As with empirical models, physical degrada- αsr
tion models can emphasize calendar aging [141] or cycle Jsr = ksr (xs,n,surf ) (92)
(1 − αsr )F αsr F
    
aging [142]. However, a better classification is to distin- × exp ηsr − exp − ηsr
guish models that focus on chemical side-reactions [129], RT RT
[141]–[145] or material fatigue [112], [131]. Rather than (93)
duplicating a review of all the models available, the rest of this ∂Ll.p.
= Cl.p. | min(ηsr , 0)| (94)
section analyzes the narrower intersection between physical ∂t
degradation modeling and optimal control.
In lithium-ion batteries, which are the primary focus of where ηsr is the side-reaction overpotential, φ1,n is the solid-
research on degradation mechanisms, the formation of the phase potential, φ2,n is the solution-phase potential, ηn is
solid electrolyte interphase (SEI) layer both increases resis- the anode overpotential, calculated using the Butler-Volmer
tance and reduces the available lithium resulting in both equation (37), 8sr is the side reaction reference voltage,
power and capacity fade [110]. The current density of the which can be conservatively estimated to be zero in this case
side-reaction that leads to the growth of the SEI layer [81], [111], Jsr is the side-reaction current density, Rfilm is the
[46], [127], [128] is shown in: lithium metal film resistance, ksr is the side-reaction rate con-
stant xs,n,surf is the surface concentration of lithium divided by
exp − RFT ηn


the maximum concentration, αsr is the side-reaction transfer
JSEI = δSEI
(88)
1  − coefficient, Cl.p. is the ratio between negative magnitude of
nSEI F nSEI F DSEI
nSEI FkSEI exp RT (8n −0.4) ηsr and the quantity of lithium-plaiting, Ll.p. is a quantitative
where JSEI is the SEI side-reaction current density, F is measure of the accumulated lithium plaiting, and Ll.p.,EoL is
Faraday’s constant, R is the ideal gas constant, T is the battery the lithium-plaiting limit at EoL. Authors in [81] simplify this
temperature, ηn is the negative electrode overpotential, nSEI by assuming that Rfilm is zero, meaning that ηsr = ηn − 8n .
is the number of electrons in the SEI side-reaction, kSEI is the Alternatively, a controller can be configured to prevent this
chemical rate constant of the SEI side-reaction, 8n is the open side-reaction entirely by constraining ηsr to be non-negative
circuit voltage of the negative electrode, δSEI is the thickness as shown in:
of the SEI layer, and DSEI is the diffusion coefficient of
lithium in the SEI layer. ηsr ≥ 0 (95)
If power is the critical parameter for operation, then we use
the growth in the thickness of the SEI layer to calculate the A controller can be designed to minimize intercalation-
rate of change in SoH as shown in: induced fatigue [112], [113]. Intercalation-induced fatigue
occurs in many battery chemistries, including lithium-ion,
∂δSEI JSEI MSEI and so this mechanism is more general than the side-reactions
= (89a)
∂t nSEI FρSEI discussed above. The radial and tangential intercalation

VOLUME 7, 2019 178383

 D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

stresses in a spherical partial can be calculated as shown in: TABLE 13. Example battery system degradation model parameters.

3 σr (r)
Z 1
1 r
Z 
2 2 2
 = xs,n (r)r dr − 3 xs,n (r)r dr
n En cn,max

3 0 r 0
(1−vpoi )
(96a)
3 σt (r) 1
Z
=2 xs,n (r)r 2 dr
n En cn,max
 
0
(1−vpoi )
1
Z r 
+ xs,n (r)r 2 dr − xs,n (r) (96b)
r3 0
∂Lstress
= max {σr (r), σt (r)} (96c)
∂t r∈[0,r̄n ]
Lstress
%stress = 1 − (96d)
Lstress,EoL
where σr is the radial intercalation stress, σt is the tangential
intercalation stress, cn,max is the maximum concentration of
lithium in the negative electrode, n is the partial molar
fd = St Sς ST + Sδ Sς ST (98c)
volume, En is Young’s modulus, vpoi is Poisson’s ratio, xs,n
is the normalized concentration in the negative electrode, St = kt n 1t (98d)
||ς||1
r is the radial distance, Lstress is the accumulated stress, and Sς = ekς ( n −ςref ) (98e)
Lstress,EoL is the accumulated stress limit at EoL. Tref
kT (||T||1 −Tref ) ||T||
It can be difficult to know how these physical degradation ST = e 1 (98f)
mechanisms combine. Each have been shown to be accurate δ = max(ς) − min(ς) (98g)
on their own, meaning that simply adding them would over- Sδ = a δ 4 + b δ 3 + c δ 2 + d δ + e (98h)
estimate the rate of degradation. One method is to calculate a
weighted combination of degradation factors as shown in: where x H = {p, pdc , ibat , vbat , voc , ς, τ, p HVAC , T, T EN ,
α% %̇stress + β% %̇stress + γ% %̇stress %̇, fd , St , Sς , ST , δ, Sδ , ∈ R9n+12 , %̇ ∈ R is the rate of
%̇ = (97) degradation, fd ∈ R is the degradation forcing function,
α% + β% + γ%
St ∈ R is the time stress-factor, Sς ∈ R is the SoC stress-
where α% , β% , and γ% are unitless weights selected to lin- factor, ST ∈ R is the temperature stress-factor, δ ∈ R is
early combine physical degradation mechanisms. However, the cycle depth-of-discharge (DoD), and Sδ ∈ R is the DoD
degradation clearly does not follow simple superposition stress-factor.
(e.g. intercalation stress and loss of lithium may have com- The net load achieved using the combined SoC-Thermal-
pounding effects) so this simplistic combination may be inac- Degradation model is shown in Fig. 27 (a). The optimal
curate. We are not aware of any experimental methods for control schedule calculated over the control horizon is shown
isolating the effects of different physical degradation mech- in Fig. 27 (b). Observe that the period of high HVAC power
anisms and so selecting weights to combine them may be in Fig. 27 (d), compared to the solution using only the SoC-
misleading. Temperature model, simply shifts to the beginning of the
control horizon. The resulting environmental, battery, and
C. DEGRADATION MODEL APPLICATION enclosure temperature trajectories are shown in Fig. 27 (e).
For this application we solve the optimal control problem in The control solution reduces the total electrical bill
Section II while incorporating the stress-factor degradation from $52,080 ($50,000 demand, $2,080 energy) to $48,006
model with parameters listed in Table 13. ($45,871 demand, $2,135 energy). Within the energy bill,
The rainflow, static-cycle model is the most appropriate the energy required to cool the battery accounts for $56. The
for this problem given that there is one-cycle that takes the net effect is a $4074 (7.82%) reduction from the baseline
whole day, and a low time resolution so low computational electrical bill, or a $5 (0.01%) increase in the electri-
burden. The modified objective and constraints, in addition cal bill calculated using only the charge and temperature
to those for the CRM and temperature models defined in models. The cost of degradation was reduced from $209,
Sections III-B.1 and IV-C.1, are shown in: calculated by applying the degradation model to the sched-
min 1t w> (l + p + p HVAC ) + ντ + C EoL %̇ (98a) ule derived from the SoC-Temperature model application
x H ∈R9n+12 in Section IV-C.1, to $111 from this schedule (a 47% reduc-
subject to: tion in estimated degradation rate). Further, when compared
.. to the solution calculated using only the SoC model in
. in addition to the constraints in (28) and (55) Section III-B.1, the cost of degradation was reduced from
%̇ = −kt Sς ST e−fd (98b) $897 to $111 (an 88% reduction). Again, these results are

178384 VOLUME 7, 2019

D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

model or a partial concentration model. The CRM tends to

simplify or combine many of the nuances of SPM type mod-
els, such as only using a single open-circuit-voltage function
as opposed to one for the cathode and another for the anode.
The SPM has the advantage of the ability to incorporate the
physical degradation models, while the ERM and CRM must
rely on empirical degradation models. To give a sense for the
relative complexity of the model types, the minimal ERM
has 3n + 2 decision variables (were n is the number of time-
steps), the minimal CRM has 6n + 2, and the minimal SPM
has 18n + 13. Further, the ERM is convex, while the CRM
and SPM have several non-convex constraints.
In Section IV we introduce three cell temperature models
and two enclosure temperature models. The primary differ-
ence between the cell models is what heating and cooling
mechanisms are considered, with another distinction being
how many internal volumes are modeled. The enclosure
models generally rely on an assumption of well-mixed air
but can be classified based on how precisely they represent
the HVAC system. As was demonstrated in the application
section, by including the HVAC system in the control design
the batteries can be pre-cooled to have maximum temperature
margin during peak discharge when significant heat is being
generated. Not accounting for temperature in control actions
can lead to over-temperature shutdown or curtailment during
peak times when the battery is needed most.
FIGURE 27. Results from control incorporating a degradation model: Lastly, Section V introduces several models to incorporate
(a) net load with BESS power control, (b) battery power, (c) battery SoC, battery degradation into control decisions. Including degra-
(d) enclosure ac power, and (e) battery, enclosure, and environmental
temperatures. dation allows for charge/discharge to be balanced against
how much the increase in use also accelerates degradation.
A wide range of empirical degradation models is available
that can be used on their own or in combinations to consider
highly conditional based on the specific BESS parameters.
many different underlying mechanisms. Physical degradation
This analysis demonstrates that even small changes in control
mechanisms are less widely used but offer the potential to
actions can have large impacts on the rate of degradation.
reduce the uncertainty of degradation modeling. Also, it can
be observed from Fig. 27 that modeling the HVAC system
VI. DISCUSSION and degradation together can have compounding benefits
Section III introduced several varieties of models for BESS to prolonging battery life. Not accounting for degradation
SoC, with the primary classification distinguishing energy in control design allows batteries to operate in ways that
based models, charge based models, and concentration-based could lead to premature EoL conditions. The following is a
models. By applying all three to the same problem we discussion of the gaps identified in the current state-of-the-
can identify several differences in how controllers might art in models for optimal control of battery energy storage.
work differently when operating with each type. Contrasting
Fig. 7 and 11, we can observe that by accounting for the
A. GAP IDENTIFICATION
change in battery voltage, the CRM steadily increases power
as SoC increases. The ERM does not model voltage and hence This paper has focused on providing guidance for how and
is imprecise in its estimates for how much power is needed to where to use different types of battery models for optimal
charge or available for discharge. In general, the ERM is best control. In this section we take a broader perspective to
for use in large scale systems where a more detailed model understand the state-of-the-art more generally and identify
would be impractical (e.g. centralized control of 1000’s of opportunities for advancement.
individual BESS) or in very short duration problems that are
insensitive to changes in voltage. The CRM and SPM are 1) REAL-WORLD DATA
mathematically similar in structure, in that they both require There is a significant deficit of operational performance data
empirical open-circuit-voltage functions and several internal in studies on optimal control of battery energy storage. This
storage elements in the form of either a equivalent circuit results in many of the modeling assumptions that proposed

VOLUME 7, 2019 178385

 D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

controllers are based on having gone unchallenged. This gap leads to higher temperatures. A cycling study cannot fully
emphasizes the importance of systematic data collection and isolate these variables because they are interlinked. Further,
publication in BESS demonstration projects. if battery degradation is a nonlinear function of both current
and temperature, then a stress factor model that assumes inde-
2) CONTROL OF LARGE-SCALE PARALLEL AND SERIES pendence will be inaccurate. The current direction of research
COMBINATIONS OF BATTERIES on degradation models improves accuracy with increasing
The ‘big-cell’ assumption is widely used to reduce the com- complexity, but controllers require computational efficiency
plexity of a large battery system to a manageable level. and hence can make limited use of these improved methods.
However, we do not know at what point the uncertainty of Research is needed to improve the accuracy of stress factor
cell-to-cell variations outweighs the uncertainty from other models that are simple enough to be incorporated into on-
modeling assumptions. Using a simpler model, with more board controllers.
representative cells, may yield better performance at a lower The literature intersecting battery energy storage modeling
complexity than a highly precise battery model that assumes and optimal control is primarily simulation based with very
all cells behave the same. This trade-off is poorly understood little work that includes experimental analysis or real-world
even though it could greatly impact BESS performance. application. This is a natural result of the combination of bat-
tery energy storage technologies having tremendous potential
3) RISK-AVERSE AND ROBUST CONTROL to change grid operation, and only recently coming down
While there exists a large body of experimental work quan- in cost enough to the point where demonstration projects
tifying the uncertainty of the different model types, this can proliferate. This means that there is significant academic
uncertainty is rarely incorporated into the battery controller interest while there are relatively few operational systems.
design. Even many controllers that consider the uncertainty A result of this lack of data is that there is little understand-
of renewable power, through risk-averse or robust control, fail ing of the impact of modeling assumptions on the design
to consider the uncertainty of the battery model’s, implicitly of controllers. Most of the gaps identified in the state-of-
assuming them to be deterministic systems. the-art stem from this lack of understanding. The remaining
gaps can be summarized as an underdeveloped optimiza-
4) NONLINEAR ERM tion framework. Stochastic optimization methods have been
Nonlinear system dynamics can be integrated into ERM used widely used in operation research to incorporate uncertainty
in controller design to improve model accuracy. The degree into the optimization problem. This mathematical back-
to which this improved model accuracy improves optimal ground has been underutilized in BESS controller design.
control is an under-explored branch of research. With more data will come improvements in the understand-
ing of uncertainty which can, in turn, be incorporated into
5) VOLTAGE HYSTERESIS IN CONTROL optimal control approaches to achieve risk averse or robust
The path dependence of open circuit voltage can be a large control.
contributor to error in SoC models. However, few con- Broadly speaking, the field on optimal control of BESS
troller designs consider hysteresis in their equivalent cir- is still nascent when compared to the markets and control
cuit, or solid-electrolyte interface voltage models. As these systems for thermal generation systems. The most com-
models are already nonlinear, and the optimal control prob- monly used models (ERM) are a simplistic approximation of
lems are already non-convex, adding hysteresis should have extremely complicated electrochemical systems. If we are to
minimal impact on computation time. learn from the historical course of optimization of thermal
generation, we can understand that simplistic models are
6) ENTROPY IN THERMAL MODELING normal at this stage of development. We may expect that these
The electrochemical reaction in batteries can be exother- models will become more developed and accurate as time
mic or endothermic, depending on the specific chemistry and progresses, leading to greater utilization of BESS to supply
the SoC. While this concept is well understood in battery services on the grid. Additionally, we might also expect that
simulation, it is rare in optimal control. Incorporating the the models used to optimize energy storage within markets
entropy-based heat generation and consumption into con- will be more abstract than the models used by individual
trollers could greatly reduce optimistic shortfall in many systems to optimize their operation. Navigating the balance
applications. between the applications that desire model simplicity and
applications that desire model accuracy will require ongoing
7) COMPARATIVE ANALYSIS OF EMPIRICAL research, especially given the accelerated pace of battery
DEGRADATION STRESS FACTORS energy storage technology development.
Battery degradation is a complex phenomenon to research.
Cycling studies try and isolate the stress factors that accel- VII. SUMMARY AND CONCLUSIONS
erate aging, but many of these factors either can’t be iso- The choice of what model to use is critical in the design of
lated or have nonlinear effects when combined with others. optimal controllers for any physical system. This is especially
For example, charge/discharge current generates heat and true for electrochemical energy storage as we have shown

178386 VOLUME 7, 2019

D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

the wide range of physical mechanisms that impact batteries ACKNOWLEDGMENT

during operation. Understanding the assumptions that are This article describes objective technical results and analysis.
implicit in the choice of battery models will help engineers Any subjective views or opinions that might be expressed
and researchers to improve the design of optimal controllers in the article do not necessarily represent the views of the
in BESS serving the electric grid. U.S. Department of Energy or the United States Government.
This paper thoroughly reviews battery models used for SAND2019-14663 J.
optimal control of BESS. We identify three broad types
of SoC models: energy reservoir models (ERMs), charge REFERENCES
reservoir models (CRMs), and concentration-based models [1] A. A. Akhil, G. Huff, A. B. Currier, B. C. Kaun, D. M. Rastler, S. B. Chen,
A. L. Cotter, D. T. Bradshaw, and W. D. Gauntlett, ‘‘DOE/EPRI 2013
that include both single partial models (SPMs) and pseudo electricity storage handbook in collaboration with NRECA,’’ Sandia
two-dimensional (P2D) models. ERMs are computationally Nat. Laboratories, Albuquerque, NM, USA, Tech Rep. SAND2013-
efficient and hence are more appropriate for the optimization 5131, Sep. 2013, [Online]. Available: http://www.sandia.
gov/ess/publications/SAND2013-5131.pdf
of large fleets of BESS. However, ERMs can be inaccurate [2] R. H. Byrne, T. A. Nguyen, D. A. Copp, B. R. Chalamala, and
over wide operational ranges. CRMs, in contrast, are more I. Gyuk, ‘‘Energy management and optimization methods for grid
computationally intensive but have the potential for better energy storage systems,’’ IEEE Access, vol. 6, pp. 13231–13260, 2018,
doi: 10.1109/ACCESS.2017.2741578.
accuracy. An under-explored middle ground for these is a [3] D. N. T. How, M. A. Hannan, M. S. H. Lipu, and P. J. Ker, ‘‘State of charge
nonlinear ERM, that may provide improved accuracy with a estimation for lithium-ion batteries using model-based and data-driven
modest increase in computational complexity. Concentration methods: A review,’’ IEEE Access, vol. 7, pp. 136116–136136, 2019,
doi: 10.1109/ACCESS.2019.2942213.
based models are significantly more complex than either [4] B. Hredzak, V. G. Agelidis, and M. Jang, ‘‘A model predictive
ERM or CRM and include many parameters that may be control system for a hybrid battery-ultracapacitor power source,’’
considered proprietary by a battery manufacturer. As the IEEE Trans. Power Electron., vol. 29, no. 3, pp. 1469–1479,
Mar. 2014, doi: 10.1109/TPEL.2013.2262003.
concentration-based SPM has a similar mathematical struc- [5] M. A. Xavier and M. S. Trimboli, ‘‘Lithium-ion battery cell-level
ture to the CRM, it is unclear how much this increase in com- control using constrained model predictive control and equivalent
plexity yields increased predictive accuracy. The SPM has circuit models,’’ J. Power Sour., vol. 285, pp. 374–384, Jul. 2015,
doi: 10.1016/j.jpowsour.2015.03.074.
the distinct advantage of enabling the application of physical
[6] J. Donadee and J. Wang, ‘‘AGC signal modeling for energy storage
degradation models that may reduce modeling uncertainty. operations,’’ IEEE Trans. Power Syst., vol. 29, no. 5, pp. 2567–2568,
When battery temperature can limit BESS control actions, Sep. 2014, doi: 10.1109/tpwrs.2014.2301592.
it is important to include a temperature model in the con- [7] E. Perez, H. Beltran, N. Aparicio, and P. Rodriguez, ‘‘Predictive power
control for PV plants with energy storage,’’ IEEE Trans. Sustain. Energy,
troller. Heat is generated in a battery from joule heating, vol. 4, no. 2, pp. 482–490, Apr. 2013, doi: 10.1109/TSTE.2012.2210255.
over potential heating, and thermodynamic entropy. While [8] H. S. Ko and J. Jatskevich, ‘‘Power quality control of wind-
the impact of entropy can be significant, it is rarely calculated hybrid power generation system using fuzzy-LQR controller,’’ IEEE
Trans. Energy Convers., vol. 22, no. 2, pp. 516–527, Jun. 2007,
in control systems. Heat is transferred from the battery, or the doi: 10.1109/TEC.2005.858092.
battery’s enclosure, to the environment through conduction, [9] D. Rosewater, Q. Nguyen, and S. Santoso, ‘‘Optimal field voltage and
convection, or radiation. Representing these mechanisms in energy storage control for stabilizing synchronous generators on flexible
AC transmission systems,’’ in Proc. IEEE/PES Transmiss. Distrib. Conf.
the controller model enables optimal cooling control that Expo. (TD), Apr. 2018, pp. 1–9, doi: 10.1109/TDC.2018.8440436.
can efficiently enforce temperature limits and significantly [10] R. Klein, N. A. Chaturvedi, J. Christensen, J. Ahmed, R. Findeisen, and
reduce battery degradation. A. Kojic, ‘‘Electrochemical model based observer design for a lithium-ion
battery,’’ IEEE Trans. Control Syst. Technol., vol. 21, no. 2, pp. 289–301,
Lastly, battery degradation can be critical to consider when Mar. 2013, doi: 10.1109/TCST.2011.2178604.
making control decisions. Factors such as SoC, temperature, [11] M. T. Lawder, B. Suthar, P. W. C. Northrop, S. De, C. M. Hoff,
and DoD can stress the materials in batteries and cause degra- O. Leitermann, M. L. Crow, S. Santhanagopalan, and V. R. Subrama-
nian, ‘‘Battery energy storage system (BESS) and battery management
dation over time and use. The empirical rainflow cycle count- system (BMS) for grid-scale applications,’’ Proc. IEEE, vol. 102, no. 6,
ing algorithm or physical degradation models for calculating pp. 1014–1030, Jun. 2014, doi: 10.1109/JPROC.2014.2317451.
degradation are both highly accurate but difficult to fully [12] V. Ramadesigan, P. W. C. Northrop, S. De, S. Santhanagopalan,
R. D. Braatz, and V. R. Subramanian, ‘‘Modeling and simula-
incorporate into an optimization problem. With a few sim- tion of lithium-ion batteries from a systems engineering perspec-
plifying assumptions, the stress factor model can be reduced tive,’’ J. Electrochem. Soc., vol. 159, no. 3, pp. R31–R45, 2012,
to take the form of a sum of regularization terms in the objec- doi: 10.1149/2.018203jes.
[13] T. A. Nguyen and R. H. Byrne, ‘‘Maximizing the cost-savings for time-
tive function. These assumptions partially justify the more
of-use and net-metering customers using behind-the-meter energy storage
widely used formulations for degradation based on number systems,’’ in Proc. North Amer. Power Symp. (NAPS), Sep. 2017, pp. 1–6,
of cycles, heat generation, and DoD. However, by including doi: 10.1109/NAPS.2017.8107380.
the nonlinear models in the controller, we observe that small [14] J. McLaren and S. Mullendore, ‘‘Identifying potential markets
for behind-the-meter battery energy storage: A survey of U.S.
changes in controller schedules can have disproportionate demand charges,’’ Nat. Renew. Energy Lab., Denver, Co, USA,
impacts on the rate of degradation. The myriad of trade- Tech. Rep. nREL/BR-6A20-68963., Aug. 2017, [Online]. Available:
offs between model complexity and model accuracy can be https://www.nrel.gov/docs/fy17osti/68963.pdf
[15] EPRI OpenDSS Test Circuits. Ckt5 Loadshape. Accessed:
difficult to navigate, but these engineering decisions can offer Nov. 2017. [Online]. Available: http://svn.code.sf.net/p/electricdss/
significant benefits in terms of BESS controller performance. code/trunk/Distrib/EPRITestCircuits/ckt5/

VOLUME 7, 2019 178387

 D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

[16] D. Rosewater, R. Baldick, and S. Santoso, ‘‘Risk-averse model predictive [36] T. Lambert, P. Gilman, and P. Lilienthal, ‘‘Micropower system
control design for battery energy storage systems,’’ IEEE Trans. Smart modeling with HOMER,’’ in Integration Of Alternative Sources
Grid, to be published, doi: 10.1109/TSG.2019.2946130. Of Energy. Hoboken, NJ, USA: Wiley, Apr. 2006, pp. 379–418,
[17] G. L. Plett, Battery Management Systems, Volume 1: Battery Modeling, doi: 10.1002/0471755621.ch15.
J. Wang, Ed. Norwood, MA, USA: Artech House, 2015. [37] G. Zini, Green Electrical Energy Storage, New York, NY, USA:
[18] A. Y. S. Lam, K.-C. Leung, and V. O. K. Li, ‘‘Capacity estimation for McGraw-Hill, 2016. [Online]. Available: http://www.ebook.de/de/
vehicle-to-grid frequency regulation services with smart charging mech- product/26046671/gabriele_zini_green_electrical_energy_storage.html
anism,’’ IEEE Trans. Smart Grid, vol. 7, no. 1, pp. 156–166, Jan. 2016, [38] T. A. Nguyen, D. A. Copp, R. H. Byrne, and B. R. Chalamala,
doi: 10.1109/tsg.2015.2436901. ‘‘Market evaluation of energy storage systems incorporating
[19] WECC Battery Storage Dynamic Modeling Guideline, Standard, West- technology-specific nonlinear models,’’ IEEE Trans. Power Syst.,
ern Electricity Coordinating Council, Salt Lake City, UT, USA, 2016. vol. 34, no. 5, pp. 3706–3715, Sep. 2019, doi: 10.1109/TPWRS.2019.
[Online]. Available: https://www.wecc.org/Reliability/WECC Battery 2909764.
Storage Guideline updates_Bo 4-5-17 SLT 4-7-17 XX SC.docx
[39] D. A. Copp, T. A. Nguyen, and R. H. Byrne, ‘‘Adaptive model predictive
[20] H. Mohsenian-Rad, ‘‘Coordinated price-maker operation of large energy control for real-time dispatch of energy storage systems,’’ in Proc. Amer.
storage units in nodal energy markets,’’ IEEE Trans. Power Syst., vol. 31, Control Conf. (ACC), Jul. 2019, pp. 3611–3616. [Online]. Available:
no. 1, pp. 786–797, Jan. 2016, doi: 10.1109/tpwrs.2015.2411556. https://ieeexplore.ieee.org/document/8814902
[21] J. Donadee and M. D. Ilić, ‘‘Stochastic optimization of grid to vehicle
frequency regulation capacity bids,’’ IEEE Trans. Smart Grid, vol. 5, [40] S.-K. Kim, J.-Y. Kim, K.-H. Cho, and G. Byeon, ‘‘Optimal operation
no. 2, pp. 1061–1069, Mar. 2014, doi: 10.1109/tsg.2013.2290971. control for multiple BESSs of a large-scale customer under time-based
pricing,’’ IEEE Trans. Power Syst., vol. 33, no. 1, pp. 803–816, Jan. 2018,
[22] T. Erseghe, A. Zanella, and C. G. Codemo, ‘‘Optimal and compact
doi: 10.1109/TPWRS.2017.2696571.
control policies for energy storage units with single and multiple batter-
ies,’’ IEEE Trans. Smart Grid, vol. 5, no. 3, pp. 1308–1317, May 2014, [41] S. Grillo, M. Marinelli, S. Massucco, and F. Silvestro, ‘‘Optimal man-
doi: 10.1109/TSG.2014.2303824. agement strategy of a battery-based storage system to improve renewable
[23] Y. Huang, S. Mao, and R. M. Nelms, ‘‘Adaptive electricity scheduling energy integration in distribution networks,’’ IEEE Trans. Smart Grid,
in microgrids,’’ IEEE Trans. Smart Grid, vol. 5, no. 1, pp. 270–281, vol. 3, no. 2, pp. 950–958, Jun. 2012, doi: 10.1109/tsg.2012.2189984.
Jan. 2014, doi: 10.1109/TSG.2013.2282823. [42] S. Teleke, M. E. Baran, S. Bhattacharya, and A. Q. Huang, ‘‘Opti-
[24] C. Suazo-Martinez, E. Pereira-Bonvallet, R. Palma-Behnke, and mal control of battery energy storage for wind farm dispatching,’’
X.-P. Zhang, ‘‘Impacts of energy storage on short term operation IEEE Trans. Energy Convers., vol. 25, no. 3, pp. 787–794, Sep. 2010,
planning under centralized spot markets,’’ IEEE Trans. Smart Grid, doi: 10.1109/tec.2010.2041550.
vol. 5, no. 2, pp. 1110–1118, Mar. 2014, doi: 10.1109/tsg.2013.2281828. [43] X. Feng, H. B. Gooi, and S. X. Chen, ‘‘Hybrid energy stor-
[25] L. H. Macedo, J. F. Franco, M. J. Rider, and R. Romero, ‘‘Optimal age with multimode fuzzy power allocator for PV systems,’’ IEEE
operation of distribution networks considering energy storage devices,’’ Trans. Sustain. Energy, vol. 5, no. 2, pp. 389–397, Apr. 2014,
IEEE Trans. Smart Grid, vol. 6, no. 6, pp. 2825–2836, Apr. 2015, doi: doi: 10.1109/tste.2013.2290543.
10.1109/tsg.2015.2419134. [44] X. Liu, P. Wang, and P. C. Loh, ‘‘A hybrid AC/DC microgrid and
[26] R. H. Byrne, M. K. Donnelly, V. W. Loose, and D. J. Trudnowski, its coordination control,’’ IEEE Trans. Smart Grid, vol. 2, no. 2,
‘‘Methodology to determine the technical performance and value pp. 278–286, Jun. 2011, doi: 10.1109/tsg.2011.2116162.
proposition for grid-scale energy storage systems,’’ Sandia Nat. [45] H. H. Abdeltawab and Y. A.-R. I. Mohamed, ‘‘Market-oriented
Laboratories, Albuquerque, NM, USA, Tech rep. SAND2012- energy management of a hybrid wind-battery energy storage sys-
10639, 2012. [Online]. Available: https://www.sandia.gov/ess- tem via model predictive control with constraint optimizer,’’ IEEE
ssl/publications/SAND2012-10639.pdf Trans. Ind. Electron., vol. 62, no. 11, pp. 6658–6670, Nov. 2015,
[27] Y. Wen, C. Guo, H. Pandžić, and D. S. Kirschen, ‘‘Enhanced security- doi: 10.1109/TIE.2015.2435694.
constrained unit commitment with emerging utility-scale energy stor- [46] S. J. Moura, J. L. Stein, and H. K. Fathy, ‘‘Battery-health conscious
age,’’ IEEE Trans. Power Syst., vol. 31, no. 1, pp. 652–662, Jan. 2016, power management in plug-in hybrid electric vehicles via electrochem-
doi: 10.1109/tpwrs.2015.2407054. ical modeling and stochastic control,’’ IEEE Trans. Control Syst. Tech-
[28] J. Tant, F. Geth, D. Six, P. Tant, and J. Driesen, ‘‘Multiobjective nol., vol. 21, no. 3, pp. 679–694, May 2013, doi: 10.1109/TCST.2012.
battery storage to improve PV integration in residential distribution 2189773.
grids,’’ IEEE Trans. Sustain. Energy, vol. 4, no. 1, pp. 182–191, [47] Y. Liu, Z. Li, Z. Lin, K. Zhao, and Y. Zhu, ‘‘Multi-objective optimization
Jan. 2013, doi: 10.1109/TSTE.2012.2211387. of energy management strategy on hybrid energy storage system based on
[29] T. Wang, H. Kamath, and S. Willard, ‘‘Control and optimization of radau pseudospectral method,’’ IEEE Access, vol. 7, pp. 112483–112493,
grid-tied photovoltaic storage systems using model predictive control,’’ 2019, doi: 10.1109/ACCESS.2019.2935188.
IEEE Trans. Smart Grid, vol. 5, no. 2, pp. 1010–1017, Mar. 2014,
[48] M. Chen and G. A. Rincon-Mora, ‘‘Accurate electrical battery
doi: 10.1109/TSG.2013.2292525.
model capable of predicting runtime and I-V performance,’’ IEEE
[30] N. Jayasekara, M. A. S. Masoum, and P. J. Wolfs, ‘‘Optimal operation of
Trans. Energy Convers., vol. 21, no. 2, pp. 504–511, Jun. 2006,
distributed energy storage systems to improve distribution network load
doi: 10.1109/tec.2006.874229.
and generation hosting capability,’’ IEEE Trans. Sustain. Energy, vol. 7,
no. 1, pp. 250–261, Jan. 2016, doi: 10.1109/TSTE.2015.2487360. [49] A. Rahmoun and H. Biechl, ‘‘Modelling of Li-ion batteries using equiv-
[31] EPRI. (Mar. 2011). OpenDSS STORAGE Element STORAGECON- alent circuit diagrams,’’ Przeglad Elektrotechniczny, vol. 88, no. 7,
TROLLER Element. [Online]. Available: http://svn.code.sf.net/p/ pp. 152–156, Jan. 2012, [Online]. Available: http://www.red.pe.org.
electricdss/code/trunk/Distrib/Doc/OpenDSS%20STORAGE%20 pl/articles/2012/7b/40.pdf
Element.pdf [50] M. Ceraolo, ‘‘New dynamical models of lead-acid batteries,’’ IEEE
[32] P. P. Zeng, Z. Wu, X.-P. Zhang, C. Liang, and Y. Zhang, ‘‘Model predictive Trans. Power Syst., vol. 15, no. 4, pp. 1184–1190, Nov. 2000,
control for energy storage systems in a network with high penetra- doi: 10.1109/59.898088.
tion of renewable energy and limited export capacity,’’ in Proc. Power [51] H. E. Perez, X. Hu, S. Dey, and S. J. Moura, ‘‘Optimal charg-
Syst. Comput. Conf., Aug. 2014, pp. 1–7, doi: 10.1109/PSCC.2014. ing of Li-Ion batteries with coupled electro-thermal-aging dynamics,’’
7038359. IEEE Trans. Veh. Technol., vol. 66, no. 9, pp. 7761–7770, Sep. 2017,
[33] P. Malysz, S. Sirouspour, and A. Emadi, ‘‘An optimal energy storage doi: 10.1109/TVT.2017.2676044.
control strategy for grid-connected microgrids,’’ IEEE Trans. Smart Grid, [52] A. M. Gee, F. V. P. Robinson, and R. W. Dunn, ‘‘Analysis of battery
vol. 5, no. 4, pp. 1785–1796, Jul. 2014, doi: 10.1109/TSG.2014.2302396. lifetime extension in a small-scale wind-energy system using supercapac-
[34] (2018). GridLAB-D Wiki: Battery Model. [Online]. Available: itors,’’ IEEE Trans. Energy Convers., vol. 28, no. 1, pp. 24–33, Mar. 2013,
http://gridlab-d.shoutwiki.com/wiki/Battery doi: 10.1109/tec.2012.2228195.
[35] X. Qiu, T. A. Nguyen, J. D. Guggenberger, M. L. Crow, and A. C. Elmore, [53] D. K. Karthikeyan, G. Sikha, and R. E. White, ‘‘Thermodynamic
‘‘A field validated model of a vanadium redox flow battery for micro- model development for lithium intercalation electrodes,’’ J.
grids,’’ IEEE Trans. Smart Grids, vol. 5, no. 4, pp. 1592–1601, Jul. 2014, Power Sources, vol. 185, pp. 1398–1407, Dec. 2008, doi:
doi: 10.1109/TSG.2014.2310212. 10.1016/j.jpowsour.2008.07.077.

178388 VOLUME 7, 2019

D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

[54] S. Santhanagopalan, Q. Guo, P. Ramadass, and R. E. White, ‘‘Review [75] O. Tremblay and L. Dessaint, ‘‘Experimental validation of a bat-
of models for predicting the cycling performance of lithium ion tery dynamic model for EV applications,’’ World Electric Vehicle
batteries,’’ J. Power Sour., vol. 156, no. 2, pp. 620–628, 2006, J., vol. 3, pp. 289–298, May 2009, [Online]. Available: https://pdfs
doi: 10.1016/j.jpowsour.2005.05.070. .semanticscholar.org/8f16/68ffef08c83a3a69f8f3c557d04cd9a5ffc2.pdf
[55] S. Santhanagopalan, Q. Guo, and R. E. White, ‘‘Parameter estimation [76] H. U. R. Habib, S. Wang, M. R. Elkadeem, and M. F. Elmorshedy,
and model discrimination for a lithium-ion cell,’’ J. Electrochem. Soc., ‘‘Design optimization and model predictive control of a Standalone
vol. 154, no. 3, pp. A198–A206, 2007, doi: 10.1149/1.2422896. hybrid renewable energy system: A case study on a small residential
[56] M. Guo and R. White, ‘‘Thermal model for lithium ion battery pack with load in Pakistan,’’ IEEE Access, vol. 7, pp. 117369–117390, 2019, doi:
mixed parallel and series configuration,’’ J. Electrochem. Soc., vol. 158, 10.1109/ACCESS.2019.2936789.
no. 10, pp. A1166–A1176, 2011, doi: 10.1149/1.3624836. [77] D. L. King, S. Gonzalez, G. M. Galbraith, and W. E. Boyson, ‘‘Per-
[57] M. Torchio, L. Magni, R. D. Braatz, and D. M. Raimondo, ‘‘Design formance model for grid-connected photovoltaic inverters,’’ Sandia Nat.
of piecewise affine and linear time-varying model predictive control Laboratories, Albuquerque, NM, USA, Tech Rep. SAND2007-5036,
strategies for advanced battery management systems,’’ J. Electrochem. 2007, [Online]. Available: http://energy.sandia.gov/download/21044/
Soc., vol. 164, no. 4, pp. A949–A959, 2017, doi: 10.1149/2.0201706jes. [78] W. Bower, C. Whitaker, W. Erdman, M. Behnke, and M. Fitzgerald,
[58] M. Torchio, N. A. Wolff, D. M. Raimondo, L. Magni, U. Krewer, ‘‘Performance test protocol for evaluating inverters used in grid-
R. B. Gopaluni, J. A. Paulson, and R. D. Braatz, ‘‘Real-time model connected photovoltaic systems,’’ Sandia Nat. Laboratories,
predictive control for the optimal charging of a lithium-ion battery,’’ Albuquerque, NM, USA, Tech. Rep., 2004. [Online]. Available:
in Proc. Amer. Control Conf. (ACC), Jul. 2015, pp. 4536–4541, doi: https://www.gosolarcalifornia.ca.gov/equipment/documents/2004-11-
10.1109/ACC.2015.7172043. 22_Test_Protocol.pdf
[59] M. Bahramipanah, D. Torregrossa, R. Cherkaoui, and M. Paolone, [79] G. L. Plett, Battery Management Systems, Volume II: Equivalent-Circuit
‘‘A decentralized adaptive model-based real-time control for active Methods, J. Wang, Ed. Norwood, MA, USA: Artech House, 2016.
distribution networks using battery energy storage systems,’’ IEEE [80] L. McCurlie, M. Preindl, and A. Emadi, ‘‘Fast model predic-
Trans. Smart Grid, vol. 9, no. 4, pp. 3406–3418, Jul. 2018, doi: tive control for redistributive lithium-ion battery balancing,’’ IEEE
10.1109/tsg.2016.2631569. Trans. Ind. Electron., vol. 64, no. 2, pp. 1350–1357, Feb. 2017, doi:
[60] R. Baldick, Applied Optimization: Formulation and Algorithms for Engi- 10.1109/TIE.2016.2611488.
neering Systems. Cambridge, U.K.: Cambridge Univ. Press, 2006. [81] J. Liu, G. Li, and H. K. Fathy, ‘‘An extended differential flatness
[61] S. Boyd and L. Vandenberghe, Convex Optimization, 1st ed. Cam- approach for the health-conscious nonlinear model predictive control of
bridge, U.K.: Cambridge Univ. Press, 2004. [Online]. Available: lithium-ion batteries,’’ IEEE Trans. Control Syst. Technol., vol. 25, no. 5,
https://web.stanford.edu/ boyd/cvxbook/ pp. 1882–1889, Sep. 2017, doi: 10.1109/TCST.2016.2624143.
[62] W. E. Hart, C. D. Laird, J.-P. Watson, D. L. Woodruff, G. A. Hackebeil, [82] X. Han, M. Ouyang, L. Lu, and J. Li, ‘‘Simplification of physics-based
B. L. Nicholson, and J. D. Siirola, Pyomo—Optimization Modeling in electrochemical model for lithium ion battery on electric vehicle. Part I:
Python, vol. 67, 2nd ed. Cambridge, U.K.: Springer, 2017. Diffusion simplification and single particle model,’’ J. Power Sour.,
[63] W. E. Hart, J.-P. Watson, and D. L. Woodruff, ‘‘Pyomo: Modeling and vol. 278, pp. 802–813, Mar. 2015, doi: 10.1016/j.jpowsour.2014.12.101.
solving mathematical programs in python,’’ Math. Program. Comput., [83] M. Pathak, S. Kolluri, and V. R. Subramanian, ‘‘Generic model con-
vol. 3, no. 3, pp. 219–260, 2011, doi: 10.1007/s12532-011-0026-8. trol for lithium-ion batteries,’’ J. Electrochem. Soc., vol. 164, no. 6,
[64] A. Wächter and L. T. Biegler, ‘‘On the implementation of an interior- pp. A973–A986, 2017, doi: 10.1149/2.1521704jes.
point filter line-search algorithm for large-scale nonlinear program- [84] N. A. Chaturvedi, R. Klein, J. Christensen, J. Ahmed, and A. Kojic,
ming,’’ Math. Program., vol. 106, no. 1, pp. 25–57, Mar. 2006, doi: ‘‘Modeling, estimation, and control challenges for lithium-ion batter-
10.1007/s10107-004-0559-y. ies,’’ in Proc. Amer. Control Conf., Jun. /Jul. 2010, pp. 1997–2002, doi:
[65] D. Linden and T. Reddy, Ed., Handbook of Batteries, New York, NY, 10.1109/ACC.2010.5531623.
USA: McGraw-Hill, 2011. [85] G. Ning and B. N. Popov, ‘‘Cycle life modeling of lithium-ion batter-
[66] A. R. Sparacino, G. F. Reed, R. J. Kerestes, B. M. Grainger, and ies,’’ J. Electrochem. Soc., vol. 151, no. 10, pp. A1584–A1591, 2004,
Z. T. Smith, ‘‘Survey of battery energy storage systems and modeling doi: 10.1149/1.1787631.
techniques,’’ in Proc. IEEE Power Energy Soc. Gen. Meeting, Jul. 2012, [86] W. M. Deen, Analysis Of Transport Phenomena, K. E. Gubbins, Ed.
pp. 1–8, doi: 10.1109/PESGM.2012.6345071. Oxford, U.K.: Oxford Univ. Press, 1998.
[67] D. Cadar, D. Petreus, I. Ciocan, and P. Dobra, ‘‘An improvement on [87] C. Y. Wang and W. B. Gu, ‘‘Micro-macroscopic coupled modeling of
empirical modelling of the batteries,’’ in Proc. 32nd Int. Spring Seminar batteries and fuel cells: I. model development,’’ J. Electrochem. Soc.,
Electron. Technol., May 2009, pp. 1–6, doi: 10.1109/ISSE.2009.5207015. vol. 145, no. 10, pp. 3407–3417, 1998, doi: 10.1149/1.1838820.
[68] G. Wang, M. Ciobotaru, and V. G. Agelidis, ‘‘Power management [88] M. Guo, G. Sikha, and R. E. White, ‘‘Single-particle model for a lithium-
for improved dispatch of utility-scale PV plants,’’ IEEE ion cell: Thermal behavior,’’ J. Electrochem. Soc., vol. 158, no. 2,
Trans. Power Syst., vol. 31, no. 3, pp. 2297–2306, May 2016, pp. A122–A132, 2011, doi: 10.1149/1.3521314.
doi: 10.1109/TPWRS.2015.2459065. [89] X. Han, M. Ouyang, L. Lu, and J. Li, ‘‘Simplification of physics-
[69] R. Xiong, J. Cao, Q. Yu, H. He, and F. Sun, ‘‘Critical review on the battery based electrochemical model for lithium ion battery on electric vehi-
state of charge estimation methods for electric vehicles,’’ IEEE Access, cle. Part II: Pseudo-two-dimensional model simplification and state of
vol. 6, pp. 1832–1843, 2017, doi: 10.1109/ACCESS.2017.2780258. charge estimation,’’ J. Power Sour., vol. 278, pp. 814–825, Mar. 2015,
doi: 10.1016/j.jpowsour.2014.08.089.
[70] Y.-H. Kim and H.-D. Ha, ‘‘Design of interface circuits with electrical
[90] A. Jokar, B. Rajabloo, and M. Désilets, and M. Lacroix, ‘‘Review
battery models,’’ IEEE Trans. Ind. Electron., vol. 44, no. 1, pp. 81–86,
of simplified pseudo-two-dimensional models of lithium-ion bat-
Feb. 1997, doi: 10.1109/41.557502.
teries,’’ J. Power Source, vol. 327, pp. 44–55, Sep. 2016, doi:
[71] A. Hentunen, T. Lehmuspelto, and J. Suomela, ‘‘Time-domain param-
10.1016/j.jpowsour.2016.07.036.
eter extraction method for Thévenin-equivalent circuit battery models,’’
[91] T. F. Fuller, M. Doyle, and J. Newman, ‘‘Simulation and optimization of
IEEE Trans. Energy Convers., vol. 29, no. 3, pp. 558–566, Sep. 2014,
the dual lithium ion insertion cell,’’ J. Electrochem. Soc., vol. 141, no. 1,
doi: 10.1109/TEC.2014.2318205.
pp. 1–10, 1994, doi: 10.1149/1.2054684.
[72] J. F. Manwell and J. G. McGowan, ‘‘Lead acid battery storage model [92] Specification Li-ion Polymer Battery 3.7V 10000 mAh, 2nd ed. Rich-
for hybrid energy system,’’ Solar Energy, vol. 50, no. 5, pp. 399–405, mond, CA USA: AA Portable Power Group, 2004, [Online]. Available:
May 1993, doi: 10.1016/0038-092X(93)90060-2. https://www.batteryspace.com/prod-specs/3.7V9059156.pdf
[73] T. Kim and W. Qiao, ‘‘A hybrid battery model capable of capturing [93] Q. Wang, B. Jiang, B. Li, and Y. Yan, ‘‘A critical review of thermal
dynamic circuit characteristics and nonlinear capacity effects,’’ IEEE management models and solutions of lithium-ion batteries for the devel-
Trans. Energy Convers., vol. 26, no. 4, pp. 1172–1180, Dec. 2011, opment of pure electric vehicles,’’ Renew. Sustain. Energy Rev., vol. 64,
doi: 10.1109/TEC.2011.2167014. pp. 106–128, Oct. 2016, doi: 10.1016/j.rser.2016.05.033.
[74] Y. Riffonneau, S. Bacha, F. Barruel, and S. Ploix, ‘‘Optimal power [94] G.-H. Kim, K. Smith, K.-J. Lee, S. Santhanagopalan, and A. Pesaran,
flow management for grid connected PV systems with batteries,’’ ‘‘Multi-domain modeling of lithium-ion batteries encompassing multi-
IEEE Trans. Sustain. Energy, vol. 2, no. 3, pp. 309–320, Jul. 2011, physics in varied length scales,’’ J. Electrochem. Soc., vol. 158, no. 8,
doi: 10.1109/tste.2011.2114901. pp. A955–A969, 2011, doi: 10.1149/1.3597614.

VOLUME 7, 2019 178389

 D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

[95] S. Barsali and M. Ceraolo, ‘‘Dynamical models of lead-acid batteries: [115] Y. Wang, Z. Song, V. D. Angelis, and S. Srivastava, ‘‘Battery life-cycle
Implementation issues,’’ IEEE Trans. Energy Convers., vol. 17, no. 1, optimization and runtime control for commercial buildings demand side
pp. 16–23, Mar. 2002, doi: 10.1109/60.986432. management: A new york city case study,’’ Energy, vol. 165, pp. 782–791,
[96] T. I. Bø and T. A. Johansen, ‘‘Battery power smoothing control in a marine Dec. 2018, doi: 10.1016/j.energy.2018.09.117.
electric power plant using nonlinear model predictive control,’’ IEEE [116] S. Ebbesen, P. Elbert, and L. Guzzella, ‘‘Battery state-of-health
Trans. Control Syst. Technol., vol. 25, no. 4, pp. 1449–1456, Jul. 2017, perceptive energy management for hybrid electric vehicles,’’ IEEE
doi: 10.1109/TCST.2016.2601301. Trans. Veh. Technol., vol. 61, no. 7, pp. 2893–2900, Sep. 2012, doi:
[97] C. Masjosthusmann, U. Köhler, N. Decius, and U. Büker, ‘‘A vehicle 10.1109/TVT.2012.2203836.
energy management system for a battery electric vehicle,’’ in Proc. [117] C. Vagg, S. Akehurst, C. J. Brace, and L. Ash, ‘‘Stochastic dynamic
IEEE Vehicle Power Propuls. Conf., Oct. 2012, pp. 339–344, doi: programming in the real-world control of hybrid electric vehicles,’’ IEEE
10.1109/VPPC.2012.6422676. Trans. Control Syst. Technol., vol. 24, no. 3, pp. 853–866, May 2016, doi:
[98] S. Mohan, J. B. Siegel, A. G. Stefanopoulou, and R. Vasudevan, ‘‘An 10.1109/TCST.2015.2498141.
energy-optimal warm-up strategy for Li-Ion batteries and its approxima- [118] C. Vagg, C. J. Brace, S. Akehurst, and L. Ash, ‘‘Minimizing bat-
tions,’’ IEEE Trans. Control Syst. Technol., vol. 27, no. 3, pp. 1165–1180, tery stress during hybrid electric vehicle control design: Real world
Mar. 2018, doi: 10.1109/TCST.2017.2785833. considerations for model-based control development,’’ in Proc. IEEE
[99] H. E. Perez, S. Dey, X. Hu, and S. J. Moura, ‘‘Optimal charging of Li- Vehicle Power Propuls. Conf. (VPPC), Oct. 2013, pp. 1–6, doi:
Ion batteries via a single particle model with electrolyte and thermal 10.1109/VPPC.2013.6671713.
dynamics,’’ J. Electrochem. Soc., vol. 164, no. 7, pp. A1679–A1687, [119] E. Vinot, R. Trigui, and B. Jeanneret, ‘‘Optimal management of electric
2017, doi: 10.1149/2.1301707jes. vehicles with a hybrid storage system,’’ in Proc. IEEE Vehicle Power
[100] P. H. L. Notten, W. S. Kruijt, and H. J. Bergveld, ‘‘Electronic network Propuls. Conf., Sep. 2010, pp. 1–6, doi: 10.1109/VPPC.2010.5729126.
modeling of rechargeable batteries: II. the NiCd system,’’ J. Electrochem. [120] Y. Tian, A. Bera, J. Mitra, B. Chalamala, and R. H. Byrne, ‘‘Effect of
Soc., vol. 145, no. 11, Nov. 1998, doi: 10.1149/1.1838872. operating strategies on the longevity of lithium-ion battery energy storage
[101] D. Rosewater, A. Headley, F. A. Mier, and S. Santoso, ‘‘Optimal control of systems,’’ in Proc. IEEE Ind. Appl. Soc. Annu. Meeting (IAS), Sep. 2018,
a battery energy storage system with a charge-temperature-health model,’’ pp. 1–8, doi: 10.1109/IAS.2018.8544518.
in Proc. IEEE PES GM, Atlanta, GA, USA, Aug. 2019. [121] B. Xu, A. Oudalov, A. Ulbig, G. Andersson, and D. S. Kirschen,
[102] H. E. Perez, X. Hu, and S. J. Moura, ‘‘Optimal charging of batteries ‘‘Modeling of lithium-ion battery degradation for cell life assessment,’’
via a single particle model with electrolyte and thermal dynamics,’’ IEEE Trans. Smart Grid, vol. 9, no. 2, pp. 1131–1140, Mar. 2018,
in Proc. Amer. Control Conf. (ACC), Jul. 2016, pp. 4000–4005, doi: doi: 10.1109/TSG.2016.2578950.
10.1109/ACC.2016.7525538. [122] Y. Niu and S. Santoso, ‘‘Sizing and coordinating fast- and slow-response
[103] H. J. Bergveld, W. S. Kruijt, and P. H. L. Notten, Battery Management energy storage systems to mitigate hourly wind power variations,’’
Systems: Design by Modelling (Battery Modeling). Berlin, Germany: IEEE Trans. Smart Grid, vol. 9, no. 2, pp. 1107–1117, Mar. 2018,
Springer, 2005. doi: 10.1109/tsg.2016.2577549.
[104] V. Viswanathan, D. W. D. Choi, S. T. W. Xu, R. E. Williford, J.-G. Zhang, [123] X. Ke, N. Lu, and C. Jin, ‘‘Control and size energy storage sys-
J. Liu, and Z. Yang, ‘‘Effect of entropy change of lithium inter- tems for managing energy imbalance of variable generation resources,’’
calation in cathodes and anodes on Li-ion battery thermal manage- IEEE Trans. Sustain. Energy, vol. 6, no. 1, pp. 70–78, Jan. 2015,
ment,’’ J. Power Sour., vol. 195, no. 11, pp. 3720–3729, 2010, doi: doi: 10.1109/tste.2014.2355829.
10.1016/j.jpowsour.2009.11.103.
[124] B. Xu, J. Zhao, T. Zheng, E. Litvinov, and D. S. Kirschen, ‘‘Factoring
[105] T. M. Bandhauer, S. Garimella, and T. F. Fuller, ‘‘A critical review of
the cycle aging cost of batteries participating in electricity markets,’’
thermal issues in lithium-ion batteries,’’ J. Electrochem. Soc., vol. 158,
IEEE Trans. Power Syst., vol. 33, no. 2, pp. 2248–2259, Mar. 2018,
no. 3, pp. R1–R25, 2011, doi: 10.1149/1.3515880.
doi: 10.1109/TPWRS.2017.2733339.
[106] D. Berndt, ‘‘Valve-regulated lead-acid batteries,’’ J. Power Sour., no. 100,
[125] Y. Shi, B. Xu, Y. Tan, D. Kirschen, and B. Zhang, ‘‘Optimal battery
pp. 29–46, Nov. 2001, doi: 10.1016/S0378-7753(01)00881-3.
control under cycle aging mechanisms in pay for performance settings,’’
[107] T. Bergman, A. Lavine, F. Incropera, and D. Dewitt, Fundamentals
IEEE Trans. Autom. Control, vol. 64, no. 6, pp. 2324–2339, Jun. 2019,
of Heat and Mass Transfer, L. Ratts, Ed. Hoboken, NJ, USA: Wiley,
doi: 10.1109/TAC.2018.2867507.
2011.
[126] H. Bindner, T. Cronin, P. Lundsager, J. Manwell, U. Abdulwahid,
[108] N. F. Ponchaut, F. Colella, V. Somandepalli, and M. Stern, ‘‘Thermal
and I. Baring-Gould, English Lifetime Modelling Of Lead Acid
management modeling for thermal runaway avoidance in lithium-ion
Batteries. Roskilde, Denmark: Risø Nat. Lab., 2005. [Online]. Available:
batteries,’’ in Proc. SAE World Congr. Exhib., Detroit, MI, USA, 2014,
https://orbit.dtu.dk/fedora/objects/orbit:88309/datastreams/file_7710966/
doi: 10.4271/2014-01-0707.
content
[109] ENERGY STAR Program Requirements for Air Source Heat Pump
(ASHP) and Central Air Conditioner Equipment - Eligibility Crite- [127] J. M. Reniers, G. Mulder, and S. Ober-Blöbaum, and D. A. Howey,
ria, 3rd ed. Washington, DC, USA, Environmental Protection Agency. ‘‘Improving optimal control of grid-connected lithium-ion batteries
[Online]. Available: https://www.energystar.gov/sites/default/files/specs/ through more accurate battery and degradation modelling,’’ J. Power
private/Room_Air_Conditioner_Program_Requirements_Version_3.pdf Sour., vol. 379, pp. 91–102, 2018, doi: 10.1016/j.jpowsour.2018.01.004.
[110] M. B. Pinson and M. Z. Bazant, ‘‘Theory of SEI formation in recharge- [128] J. Christensen and J. Newman, ‘‘Cyclable lithium and capacity loss in Li-
able batteries: Capacity fade, accelerated aging and lifetime predic- Ion cells,’’ J. Electrochem. Soc., vol. 152, no. 4, pp. A818–A829, 2005,
tion,’’ J. Electrochem. Soc., vol. 160, no. 2, pp. A243–A250, 2013, doi: doi: 10.1149/1.1870752.
10.1149/2.044302jes. [129] M. Safari, M. Morcrette, A. Teyssot, and C. Delacourt, ‘‘Multi-
[111] P. Arora, M. Doyle, and R. E. White, ‘‘Mathematical modeling of the modal physics-based aging model for life prediction of Li-ion bat-
lithium deposition overcharge reaction in lithium-ion batteries using teries,’’ J. Electrochem. Soc., vol. 156, no. 3, pp. A145–A153, 2009,
carbon-based negative electrodes,’’ J. Electrochem. Soc., vol. 146, no. 10, doi: 10.1149/1.3043429.
pp. 3543–3553, 1999, doi: 10.1149/1.1392512. [130] M. K. S. Verma, S. Basu, K. S. Hariharan, S. M. Kolake, T. Song, and
[112] B. Suthar, P. W. C. Northrop, R. D. Braatz, and V. R. Subramanian, J. Jeon, ‘‘A strain-diffusion coupled electrochemical model for lithium-
‘‘Optimal charging profiles with minimal intercalation-induced stresses ion battery,’’ J. Electrochem. Soc., vol. 164, no. 13, pp. A3426–A3439,
for lithium-ion batteries using reformulated pseudo 2-dimensional mod- 2017, doi: 10.1149/2.0021714jes.
els,’’ J. Electrochem. Soc., vol. 161, no. 11, pp. F3144–F3155, 2014, doi: [131] M. Safari, M. Morcrette, A. Teyssot, and C. Delacourt, ‘‘Life-prediction
10.1149/2.0211411jes. methods for lithium-ion batteries derived from a fatigue approach i.
[113] Y.-T. Cheng and M. W. Verbrugge, ‘‘Evolution of stress within a introduction: Capacity-loss prediction based on damage accumulation,’’
spherical insertion electrode particle under potentiostatic and gal- J. Electrochem. Soc., vol. 157, no. 6, pp. A713–A720, 2010, doi:
vanostatic operation,’’ J. Power Sour., vol. 190, no. 2, pp. 453–460, 10.1149/1.3374634.
2009, doi: 10.1016/j.jpowsour.2009.01.021. [132] J. Schmalstieg and S. Käbitz, M. Ecker, and D. U. Sauer, ‘‘A holis-
[114] U. Car, ‘‘Battery test manual for electric vehicles,’’ Idaho Nat. Lab., Idaho tic aging model for Li(NiMnCo)O2 based 18650 lithium-ion bat-
Falls, ID, USA, Tech Rep. INL/EXT-15-34184 2015, [Online]. Available: teries,’’ J. Power Source, vol. 257, pp. 325–334, Jul. 2014, doi:
http://www.uscar.org/commands/files_download.php?files_id=405 10.1016/j.jpowsour.2014.02.012.

178390 VOLUME 7, 2019

D. M. Rosewater et al.: Battery Energy Storage Models for Optimal Control

[133] Y. Cui, C. Du, G. Yin, Y. Gao, L. Zhang, T. Guan, L. Yang, and F. Wang, DAVID A. COPP received the B.S. degree in
‘‘Multi-stress factor model for cycle lifetime prediction of lithium ion mechanical engineering from the University of
batteries with shallow-depth discharge,’’ J. Power Sources, vol. 279, Arizona and the M.S. and Ph.D. degrees in
pp. 123–132, Apr. 2015, doi: 10.1016/j.jpowsour.2015.01.003. mechanical engineering from the University of
[134] J. Wang, J. Purewal, P. Liu, J. Hicks-Garner, S. Soukazian, E. Sherman, California, Santa Barbara. He was a Senior Mem-
A. Sorenson, L. Vu, H. Tataria, and M. W. Verbrugge, ‘‘Degradation ber of the Technical Staff with the Sandia National
of lithium ion batteries employing graphite negatives and nickel-cobalt- Laboratories and an Adjunct Faculty Member in
manganese oxid + spinel manganese oxide positives: Part 1, aging electrical and computer engineering with the Uni-
mechanisms and life estimation,’’ J. Power Sour., vol. 269, pp. 937–948,
versity of New Mexico. He is currently an Assis-
Dec. 2014, doi: 10.1016/j.jpowsour.2014.07.030.
tant Professor of teaching with the Department of
[135] S. D. Downing and D. F. Socie, ‘‘Simple rainflow counting algorithms,’’
Mechanical and Aerospace Engineering, University of California, Irvine. His
Int. J. Fatigue, vol. 4, no. 1, pp. 31–40, Jan. 1982, doi: 10.1016/0142-
1123(82)90018-4. broad research interests include engineering education, as well as control and
[136] E. Thomas, I. Bloom, J. Christophersen, and V. Battaglia, ‘‘Statistical
optimization of nonlinear and hybrid systems with applications to power and
methodology for predicting the life of lithium-ion cells via accelerated energy systems, multiagent systems, robotics, and biomedicine. He was a
degradation testing,’’ J. Power Source, vol. 184, no. 1, pp. 312–317, 2008, recipient of the UCSB’s Center for Control, Dynamical Systems, and the
doi: 10.1016/j.jpowsour.2008.06.017. Computation Best Ph.D. Thesis Award.
[137] M. Ecker, J. B. Gerschler, J. Vogel, S. Kabitz, F. Hust, P. Dechent, and
D. U. Sauer, ‘‘Development of a lifetime prediction model for lithium-ion TU A. NGUYEN received the B.S. degree in
batteries based on extended accelerated aging test data,’’ J. Power Source, power systems from the Hanoi University of Sci-
vol. 215, pp. 248–257, Oct. 2012, doi: 10.1016/j.jpowsour.2012.05.012. ence and Technology, Vietnam, in 2007, and the
[138] K. Abdulla, J. de Hoog, V. Muenzel, F. Suits, K. Steer, A. Wirth, and Ph.D. degree from the Missouri University of Sci-
S. Halgamuge, ‘‘Optimal operation of energy storage systems considering ence and Technology, in December 2014. He was
forecasts and battery degradation,’’ IEEE Trans. Smart Grid, vol. 9, no. 3, worked as a Power Transformer Test Engineer with
pp. 2086–2096, May 2018, doi: 10.1109/TSG.2016.2606490. the ABB High Voltage Test Department, Vietnam,
[139] M. R. Palacin and A. de Guibert, ‘‘Why do batteries fail?’’ Science, from 2008 to 2009. He was worked as a Post-
vol. 351, no. 6273, Feb. 2016, Art. no. 1253292, doi: 10.1126/sci- doctoral Research Associate with the University
ence.1253292. of Washington. In September 2016, he joined the
[140] Y. K. Lee, J. Park, and W. Lu, ‘‘A comprehensive experimental and Sandia National Laboratories, where he is currently a Senior Member of
modeling study on dissolution in Li-Ion batteries,’’ J. Electrochem. Soc., the Technical Staff. His research interests include energy storage economics,
vol. 166, no. 8, pp. A1340–A1354, 2019, doi: 10.1149/2.0111908jes. microgrid modeling and analysis, and the integration of distributed resources
[141] T. Lu, Y. Luo, Y. Zhang, W. Luo, L. Yan, and J. Xie, ‘‘Degra- into power grids.
dation analysis of commercial lithium-ion battery in long-term stor-
age,’’ J. Electrochem. Soc., vol. 164, no. 4, pp. A775–A784, 2017, RAYMOND H. BYRNE (F’17) received the B.S.
doi: 10.1149/2.1321704jes. degree in electrical engineering from the Univer-
[142] J. Wang, P. Liu, J. Hicks-Garner, E. Sherman, S. Soukiazian, sity of Virginia, the M.S. degree in electrical engi-
M. Verbrugge, H. Tataria, J. Musser, and P. Finamore, ‘‘Cycle-life
neering from the University of Colorado, the M.S.
model for graphite-lifepo4 cells,’’ J. Power Source, vol. 196, no. 8,
degree in financial mathematics (financial engi-
pp. 3942–3948, 2011, doi: 10.1016/j.jpowsour.2010.11.134.
neering) from The University of Chicago, and the
[143] P. Arorat, R. E. White, and M. Doyle, ‘‘Capacity fade mechanisms and
side reactions in lithium-ion batteries,’’ J. Electrochem. Soc., vol. 145, Ph.D. degree in electrical engineering from The
no. 10, pp. 3647–3667, 1998, doi: 10.1149/1.1838857. University of New Mexico. He is currently a Dis-
[144] T. Waldmann, B.-I. Hogg, and M. Wohlfahrt-Mehrens, ‘‘Li tinguished Member of the Technical Staff with the
plating as unwanted side reaction in commercial Li-ion cells—A Sandia National Laboratories, where he has been
review,’’ J. Power Source, vol. 384, pp. 107–124, Apr. 2018, doi: employed, since 1989. He also serves as the Team Lead of the Equitable
10.1016/j.jpowsour.2018.02.063. Regulatory Environment thrust area of the Sandia energy storage program.
[145] P. Ramadass, B. Haran, P. M. Gomadam, R. White, and B. N. Popov, He was elevated to the IEEE Fellow for contributions to miniature robotics
‘‘Development of first principles capacity fade model for Li-Ion cells,’’ and grid integration of energy storage, in 2017. He is a member of Tau
J. Electrochem. Soc., vol. 151, no. 2, pp. A196–A203, 2004, doi: Beta Pi, Eta Kappa Nu, and Sigma Xi. He was recipient of several awards,
10.1149/1.1634273. including Time Magazine invention of the year in robotics, in 2001, and the
Prize Paper Award at the 2016 IEEE Power and Energy Society General
Meeting for a article on maximizing revenue from energy storage in grid
applications.

SURYA SANTOSO (F’15) received the B.S.

degree from Satya Wacana Christian University,
Salatiga, Indonesia, in 1992, and the M.S.E. and
Ph.D. degrees in electrical and computer engi-
neering from The University of Texas at Austin,
in 1994 and 1996, respectively. He was a Senior
DAVID M. ROSEWATER (M’11) is currently Power Systems and Consulting Engineer with
pursuing the Ph.D. degree in electrical and com- Electrotek Concepts, Knoxville, TN, USA, from
puter engineering with UT Austin. He worked 1997 to 2003. He joined the faculty of The Uni-
with the Idaho National Laboratory developing versity of Texas at Austin, in 2003, where he is
advanced spectral impedance measurement tech- currently a Professor of electrical and computer engineering. He is a coauthor
niques for hybrid vehicle battery cells. He is cur- of Electrical Power Systems Quality (third edition), sole author of Funda-
rently a Senior Member of the Technical Staff with mentals of Electric Power Quality, and an Editor of Handbook of Electric
the Sandia National Laboratories. He researches Power Calculations (fourth edition) and Standard Handbook for Electrical
applications of energy storage technologies in Engineers (17th edition). His research interests include power quality, power
electric power systems. His research interests systems, and renewable energy integration in transmission, and distribution
include modeling and simulation, performance testing, safety, and standard- systems.
ization of battery energy storage systems.

VOLUME 7, 2019 178391