Innovative Energy Arbitrage Models and Algorithms for Battery

Energy Storage Systems in Electricity Market

This paper was downloaded from TechRxiv (https://www.techrxiv.org).

LICENSE

CC BY 4.0

SUBMISSION DATE / POSTED DATE

25-10-2021 / 28-10-2021

CITATION

Li, Ang; Peng, Jiming; Fan, Lei; Du, Pengwei (2021): Innovative Energy Arbitrage Models and Algorithms for
Battery Energy Storage Systems in Electricity Market. TechRxiv. Preprint.
https://doi.org/10.36227/techrxiv.16867315.v1

DOI

10.36227/techrxiv.16867315.v1
 1

Innovative Energy Arbitrage Models and

Algorithms for Battery Energy Storage Systems in
Electricity Market
Ang Li; Jiming Peng; Lei Fan, Senior Member, IEEE; Pengwei Du, Senior Member, IEEE

Abstract—Maintaining the balance between electricity produc- Ic The maximum charging current for the battery
tion and consumption is an essential task in the operations of energy storage.
modern power grids. In recent years, battery energy storage sys- Id The maximum discharging current for the battery
tem (BESS) has been gaining more and more attention owning to
its decreasing capital cost, high flexibility and short response time. energy storage.
However, there exist several technical challenges in developing λt Electricity price at period t.
accurate models and effective algorithms for the operations of ηc The charging efficiency.
BESS. One major challenge is that in order to precisely describe ηd The discharging efficiency.
the changing dynamics of battery status, usually highly non-linear ∆ The length of time interval.
functions and integer variables must be used in the model, leading
to optimization models with nonlinear and non-convex objective B. Decision Variables
function/constraints that are difficult to handle. pct The charging power of the battery energy storage
To address the above challenges, in this paper we first at period t.
explore the physical law in the battery charging and discharging pdt The discharging power of the battery energy stor-
process to develop a new model which naturally addresses age at period t.
both the charging and discharging processes through a single
current decision variable. Then we propose to approximate the st The state of charge (SOC) of the battery energy
dependency relationship between the open circuit voltage (OCV) storage at time period t.
and the state of charge (SOC) by some linear functions. This it The charging/discharging current of the battery
leads to a new non-convex quadratic programming model with energy storage at time period t.
linear and bilinear constraints (BLCQP) for the identification vtoc The open circuit voltage of the battery energy
of the optimal operational strategy for the BESS. To cope with
the bilinear constraints in the BLCQP, we introduce a novel storage at time period t.
transformation technique to transfer the original BLCQP into C. Abbreviations
another equivalent exponential optimization problem with linear BESS Battery energy storage systems.
constraints (LCEO). A new sequential linear and quadratic OCV The open circuit voltage.
programming approach (SLQP) for the LCEO is developed SOC The state of charge.
and its convergence is established. Preliminary experiments are
conducted to demonstrate the efficacy of the new model and the OCVDF The dependency function of the open circuit volt-
efficiency of the new algorithm. age on the state of charge.
Index Terms—Battery Energy Storage, Electricity Market,
PAM The power arbitrage model.
Energy Arbitrage Model, Nonconvex Quadratic Programming. VIAM The voltage (V) current (I) arbitrage model.
BLCQP Bi-linear constrained quadratic programming.
LCEO Linear constrained optimization with exponential
N OMENCLATURE objective function.
A. Parameters SLQP The sequential linear and quadratic programming
T The set of operation time span. algorithm.
EC The energy capacity of the battery energy storage.
I. I NTRODUCTION
R The electrical resistance of the battery energy
storage. In recent years, due to climate change, extreme weather
Ptc The maximum charging rate for the battery energy events have occurred frequently. For example, the 2021 North
storage. American winter storm caused widespread blackouts in Texas,
Ptd The maximum discharging rate for the battery which led to heavy casualties and property losses. To slow the
energy storage. process of climate warming, the U.S. government has set a
goal of achieving net-zero carbon emissions by 2050. This is a
Ang Li and Jiming Peng is with the Department of Industrial Engineering, task that cannot be done without the help of renewable energy
University of Houston, Houston, TX 77204. E-mail: [email protected],
[email protected], resources. However, renewable energy resources such as wind
Lei Fan (Corresponding Author) is with the Department of Engi- energy and solar energy are changing the operational mode
neering Technology, University of Houston, Houston, TX 77204. Email: of the power system because traditional decision methods and
[email protected].
Pengwei Du is with Electric Reliability Council of Texas, Austin, TX. approaches cannot deal with the high volatility and uncertainty
Email: [email protected] of these renewable energy resources [1].
 2

In order to maintain the reliability and flexibility of the physical principles in the process of battery charging and
power system under the high penetration rate of renewable discharging, we propose to use only a single current variable
energy, highly flexible resources such as energy storage have (it ) to model both the charging and discharging processes in
become one of the important resources for the system to the BESS. Specifically, it > 0 indicates that the battery is in
respond to and track the dynamic changes caused by wind the charging process and it < 0 indicates that the battery is
and solar energy [2]. Among different types of energy storage in discharging process. Consequentially, such a formulation
resources such as pump energy storage and compressed air ensures that both charging and discharging won’t happen
energy storage, battery energy storage has become a popular simultaneously, which provides a natural way to model the
choice due to its wide range of advantages. For example, charging/discharging process of the battery. Mathematically
unlike pump energy storage plants and compressed air energy speaking, this also allows us to get rid of the complimentary
storage, which are restricted by geographic locations, battery conditions (i.e., pct pdt = 0) in most existing optimization
energy storage can be installed not only in power generation models in the literature and thus help to substantially simplify
systems, but also in power distribution systems. In addition, these models.
the cost of battery energy storage has dropped significantly To address the second challenge, we observe that one
according to [3]. major nonlinearity and non-convexity stem from the OCVDF
Battery energy storage resources have been broadly studied when the detailed BESS optimization model is applied [19].
in many areas including power system planning [4], power However, as pointed out by [21], normally the state of charge
system operations [5], power distribution system [6], bidding is operated in the range [20%, 80%] because the lifetime of
strategy [7], virtual power plants [8], and micro-grid systems the battery will be reduced significantly when it is operated
[9]. In these applications, most researchers use a high level outside such a range. When the SOC is restricted to the range
optimization model to describe the battery energy storage of [20%, 80%], we can use some linear functions to obtain
operational process, in which the charging and discharging a very good approximation to the original nonlinear OCVDF
processes are controlled by two different variables [10]. Be- (g(s)). This can be verified from our experiment. For example,
cause the charging and discharging can not happen at the if we use the OCVDF defined in [22] as the original OCVDF
same time, some complementarity conditions are imposed to and apply the least squares method in Banach space [23]
ensure that at most one of these two variables can take nonzero to obtain the linear approximate function, then we need to
values. From a mathematical perspective, the complementary solve the following optimization problem to estimate the two
conditions can also be modeled using bilinear constraints [11] parameters c0 and c1 in the approximate linear function.
or binary variables [12]. The presence of massive bilinear
constraints has become a major challenge in the development Z 0.8

of effective algorithms for these models. Researchers [13] min L(ḡ) = (ḡ(s) − g(s))2 ds.
su ,ḡ 0.2
investigated the possibility to get rid of the bilinear constraints
s.t. ḡ(s) = c0 + c1 s.
and binary variables to have a convex optimization model.
However, as pointed by [10], these convex relaxation model
will result in impractical solutions which might be infeasible. By solving the above convex optimization problem, we obtain
Recently, researchers [14] and [15] studied more detailed the approximation function ḡ(s) = 14.8629s + 126.196. The
battery energy storage model with considering the dependency following figure shows the original OCVDF g(s) and its linear
relationship between the open circuit voltage (OCV) and SOC. approximation ḡ(s).
The open-circuit voltage voc , representing the voltage differ-
ence between the electrodes at the open circuit, depends on the
SOC. In the remaining parts of this paper, for convenience, we
assume the dependency relationship is defined by the so-called
open-circuit voltage dependency function (OCVDF). Various
OCVDF formulations have been proposed in the literature and
most of these formulations are highly non-linear and non-
convex [14], [16]–[20], which further leads to non-convex
optimization models with non-convex constraints for the BESS
that are difficult to solve. This forms the second major compu-
tational challenge in the BESS optimization models. Although
[14] and [15] provide a new detailed battery energy storage
model, the authors ignored the complementary conditions and Fig. 1: The OCVDF and its linear approximation
did not discuss how to cope with the nonlinearity and non-
convexity arisen from the OCVDF. To estimate how the linear function approximates the
The main goal of this work is to address the above- OCVDF, we compute the maximal relative error via solving
mentioned major challenges in the BESS optimization models. the following optimization problem:
To address the first major challenge, we will review the basic
BESS optimization model widely used in the literature using |ḡ(s) − g(s)|
two charging/discharging power variables. By exploring the ρ= max .
s∈[0.2,0.8] g(s)
 3

In the experiment, we obtain ρ = 0.597462%. It follows state of charge balance (1f). This PAM has been widely used in
different power system applications such as unit commitment,
ḡ(s)
0.994 ≤ 1 − ρ ≤ ≤ 1 + ρ ≤ 1.006, ∀s ∈ [0.2, 0.8]. economic dispatch, transmission planning, etc. However, this
g(s)
model ignores the relationship between the state of charge
The above relation demonstrates that though the linear ap- and voltage of the storage and thus significantly simplifies the
proximation may lose some accuracy at certain points, it does battery energy storage physical operation process. As pointed
provide a very good approximation to the original OCVDF as out by Arroyo et al. in [10], this simplified model may lead
a whole. to infeasible dispatch strategy for the field operations of the
As we shall see in our later analysis, even when the OCVDF BESS.
is linear, the resulting BESS optimization model reduces to
a non-convex quadratic programming with linear and bilinear A. Physical Model
constraints (BLCQP), which is still hard to solve. The bilinear
term results from the product of the voltage and current. In this section, we propose a brand new BESS schedule
To cope with the bilinear constraints in the corresponding model based on equivalent circuit model by using the funda-
BESS optimization model, we will introduce a novel nonlinear mental physical laws. In the following, we always assume the
transformation to reformulate the original BLCQP into another time span we study is T = {1, 2, 3, . . . , T }.
equivalent exponential optimization problem with only linear
constraints (LCEO). We then modify the standard sequential
quadratic programming approach to develop the so-called
sequential linear/quadratic programming (SLQP) for the re-
formulated LCEO model and establish the convergence for
the proposed SLQP. We also conduct numerical experiments
to verify the efficacy of our new BESS optimization model
and the efficiency of the new SLQP algorithm.
Fig. 2: Equivalent Circuit
This paper is organized as follows. In Section II, we first
describe the basic BESS optimization model in the literature Figure 2 shows the circuit of BESS. Let R be the equivalent
and then explore the physical laws in the charging/discharging resistance. When BESS is in the discharging status, we have
process to simplify the basic model. In Section III, we intro- the following equations.
duce a nonlinear transformation technique to reformulate the
original BLCQP as the equivalent LCEO. In Section IV, we vtoc,d = vtout,d + idt R;
describe the SLQP approach for the reformulated LCEO model pdt = vtout,d idt .
and establish the convergence of the proposed algorithm. In
Section V, we present a case study using electricity price and When BESS is in the charging status, we have the following
battery data to demonstrate the efficacy of our new models equations.
and techniques. In section VI, we conclude the paper by
vtoc,c = vtout,c − ict R;
summarizing the topics discussed.
pct = vtout,c ict .
II. BATTERY A RBITRAGE O PTIMIZATION M ODEL IN Now, we use general formulation to characterize both the
E LECTRICITY M ARKET charging and discharging status. The positive value of current
In the literature [10], [14], [16], the essential part of BESS i represents that the battery is charging. Then negative value
power arbitrage model (PAM) can be extracted as follows: of current i represents that the battery is discharging. Then we
X have.
(P AM ) min λt (pct − pdt ) (1a)
pd c
t ,pt ,st t∈T vtoc = vtout − it R; (2a)
s.t. pct pdt = 0, ∀t ∈ T ; (1b) pout
t = vtout it . (2b)
c
0≤p ≤ Ptc , ∀t ∈ T ; (1c) Using this general formulation, we may express the objec-
tive function (1a) in PAM as t∈T λt pout = t∈T λt (vtoc it +
d
P P
0≤p ≤ Ptd , ∀t ∈ T ; (1d) t

sl ≤ st ≤ s , ∀t ∈ T ;
u
(1e) i2t R), and eliminate the complimentary constraint (1b). By us-
ing the equivalent circuit model (2a) - (2b), we can reformulate
pd the PAM as the voltage (V) current (I) arbitrage model (VIAM)
ECst+1 = ECst +pct η c ∆− td ∆, ∀t ∈T. (1f)
η as follows.
In the electricity market, the BESS owners use the PAM to X
(V IAM ) ocmin λt (vtoc it + i2t R) (3a)
help them make decisions. In this PAM, the objective function vt ,it ,st
t∈T
is to minimize the costs (or maximize the profits equivalently)
s.t. ECst+1 = ECst + vtoc it ∆, ∀t ∈ T ; (3b)
of the BESS in the electricity market. This PAM includes the
complementary constraints (1b), the charging/discharging rates vtoc = g(st ), ∀t ∈ T ; (3c)
limits (1c-1d), the operational ranges constraints (1e) and the s1 = sT +1 = s̄; (3d)
 4

sl ≤ st ≤ su , ∀t ∈ T ; (3e) Then we obtain the following quadratic programming model

− I d ≤ it ≤ I c , ∀t ∈ T . (3f) with linear and bi-linear constraints (BLCQP):
X
Notice in this VIAM model, we consider the dependency (BLCQP ) min λt (gt it + i2t R) (4a)
it ,gt
t∈T
of OCV on SOC by constraint (3c). We also replace the
limits of charging/discharging power rates with the limits of s.t. gt+1 = gt + τ gt it , ∀t ∈ T ; (4b)
charging/discharging current rates, this corresponds to con- g1 = gT +1 = ḡ; (4c)
straint (3f). We note in VIAM model, in order to compare l u
g ≤ gt ≤ g , ∀t ∈ T ; (4d)
the solution quality, we assume that the SOC levels s equal d c
at the beginning and the end of studied time interval T , i.e., − I ≤ it ≤ I , ∀t ∈ T . (4e)
s1 = sT +1 . A specific choice is s1 = sT +1 = s̄ = 0.5. Under We have
such an assumption, the profit comes only from the trading of
electricity. Theorem 1. Model (4) is equivalent to model (3).
As shown in the survey papers [17]–[20] and the refer- Proof. The sufficiency follows from our deduction of the
ences therein, the OCVDF g(s) is usually highly nonlinear BLCQP model and thus it remains to prove the necessity.
and non-convex. The non-linearity and non-convexity in both Suppose that gt is a feasible solution to model (4a). Since
the constraint functions and the objective function pose a g(s) is a linear bijection, its inverse function g −1 exists and is
tremendous computational challenge to even obtain a local also linear. By applying g −1 to both sides of constraint (4b),
optimal solution. To tackle such a challenge, we propose we obtain
to approximate the OCVDF g(s) by some relatively simple τ
functions. For example, if we assume the OCVDF g(s) is g −1 (gt+1 ) = g −1 (gt + τ gt it ) = g −1 (gt ) + gt it ,
c1
a constant, then the above model can be reduced to convex
quadratic optimization which can be solved by many off-shelf which is equivalent to constraint (3b). Therefore, we obtain a
optimization software such as [24]. In this paper, we will solution st = g −1 (gt ), ∀t ∈ T to model (3a). This completes
consider the scenario where the OCVDF g(s) is approximated the proof of the theorem.
by a linear function and discuss how to solve the resulting Note that in model (4a), there are some non-convex
bilinear optimization model in the following section. quadratic (or bilinear) terms in both objective function and
constraint (4b). This implies that it is still nontrivial to solve
model (4a) due to the presence of the non-convex quadratic
III. A N ON - CONVEX Q UADRATIC P ROGRAMMING M ODEL
functions.
AND R EFORMULATION T ECHNIQUES

In this section, we consider a special case of VIAM model B. A Nonlinear Transformation

In this subsection we introduce a nonlinear transformation
(3a) where the OCVDF g(s) is an linear function. This leads
to simplify the bilinear constraints in model (4a). For this, we
to a non-convex quadratic programming model with linear
note that in the constraint (4b), if we take the logarithm on
and bi-linear constraints (BLCQP). We then introduce a novel
both sides, we obtain:
nonlinear transformation to reformulate the original BLCQP
model as another equivalent optimization problem with linear log(gt+1 ) = log(gt ) + log(1 + τ it )
constraints and exponential objective function (LCEO).
Let
A. A Quadratic Programming Model Based on Linear Approx-
imation of OCVDF yt = log(gt ), zt = log(1 + τ it ), ∀t ∈ T , (5)
In this subsection, we discuss how to solve model (3a)
we can rewrite the constraint (4b) as
where the OCVDF is linear or approximated by an linear
function. Let g(s) = c0 + c1 s denote the linear approximate yt+1 = yt + zt , ∀t ∈ T .
function. By applying the relation defined by constraint (3b)
to the function g(s), we obtain a new constraint as follows Moreover, constraint (4c) can be rewritten as y1 = yT +1
zt
equivalently. Recall that gt = eyt and it = e τ−1 , we
∆ oc reformulate the objective function by using constraint (4b),
g(st+1 ) = g(st + v it )
EC t and new variables yt , zt ,
∆ oc
= g(st ) + c1 v it gt+1 − gt
EC t λt (gt it + i2t R) = λt ( + i2t R)
∆ τ
= g(st ) + c1 g(st )it . λt λt R
EC = (eyt+1 − eyt ) + 2 (ezt − 1)2
τ τ
Let g l = c0 + c1 sl and g u = c0 + c1 su be the lower and
Let
upper bound for the linear OCVDF g(s), then we can replace
the box constraint (3e) on SOC by that of OCVDF g(s). For ȳ = log(g1 ) = log(gT +1 ) = log(ḡ);
notational convenience, let y l = log(g l ), y u = log(g u );
∆ z l = log(1 − τ I d ), z u = log(1 + τ I c ).
τ = c1 , ḡ = g(s̄), gt = g(st ), ∀t ∈ T .
EC
 5

Using the new variables yt and zt as defined in (5), we can To find a search direction (dky , dkz ), we propose to solve the
recast model (4a) as the following optimization problem with following convex quadratic optimization problem
linear constraints (LCEO): X
X λt−1 − λt min f¯1 (ytk + dt,k ¯ k t,k
y ) + f2 (zt + dz ) (7a)
λt R dt,k t,k
y ,dz
(LCEO) min eyt + 2 (ezt − 1)2 (6a) t∈T
yt ,zt τ τ
t∈T s.t. dt+1,k
y = dt,k t,k
y + dz , ∀t ∈ T ; (7b)
s.t. yt+1 = yt + zt , ∀t ∈ T ; (6b) d1,k
y = 0; (7c)
y1 = ȳ; (6c) dTy +1,k = 0; (7d)
yT +1 = ȳ; (6d) y l ≤ ytk + dt,k u
y ≤ y , ∀t ∈ T ; (7e)
y l ≤ yt ≤ y u , ∀t ∈ T ; (6e) z l ≤ ztk + dt,k u
z ≤ z , ∀t ∈ T . (7f)
z l ≤ zt ≤ z u , ∀t ∈ T . (6f)
Theoretically, the above problem is polynomially solvable
We remark that comparing with the original bilinear optimiza- [25]. In our experiments, we use the optimization software
tion model (4), model (6) has only linear constraints with an CVX [24] and IPOPT [26] to solve the above problem. Once
exponential objective function. Notice that, if a search direction is identified, we need to find a step size α
1 satisfying the following relation to ensure sufficient descent:
λt−1 ≥ λt , ezt ≥
, ∀t ∈ T ,
2 X
f1 (ytk + αdk,t k t,k
y ) + f2 (zt + αdz ) (8)
then model (6) becomes a convex optimization problem. t∈T
Particularly, because ezt ≥ ezl = 1−τ I d , the relation ezt ≥ 21 X α 0 k t,k
is usually satisfied in our problem setting due to the fact that ≤ f1 (ytk ) + f2 (ztk ) + (f (y )d + f20 (ztk )dt,k
z ).
2 1 t y
t∈T
τ I d is much smaller than 1. It also worths mentioning that
when the prices remain invariant, i.e., λ1 = λ2 = · · · = λT , Because the derivatives of f¯i and fi always coincide with
then model (6) has a trivial optimal solution (y1 = y2 = · · · = each other, we do not distinguish f¯i0 and fi0 in inequality (8)
yT , z1 = z2 = · · · = zT = 0). This implies that a flat price and in the remaining parts of this work. Particularly, we
dynamic will lead to less profit in the BESS. adapt the Armijo–Goldstein line search procedure [27] for
nonlinear optimization to find a suitable step size α as shown
IV. A S EQUENTIAL L INEAR /Q UADRATIC P ROGRAMMING in algorithm 1.
A PPROACH
In this section, we propose a sequential linear/quadratic Algorithm 1 Line Search Procedure (LSP)
programming approach (SLQP) for the reformulated LCEO Input: f1 , f2 , f10 , f20 , y k , z k , dky , dkz ,  > 0, 0 < η < 1.
model introduced in the previous section and establish its Output: Stepsize αk .
convergence. Begin
To start, we mention that the solution (y (0) , z (0) ) = (y1 = Set α = αmax .
y2 = · · · = yT +1 , z1 = z2 = · · · = zT = 0) is feasible for the while f1 (y k +αdky )+f2 (z k +αdkz ) > f1 (y k )+f2 (z k )+
LCEO and it strictly satisfies the two box constraints. α 0 k k 0 k k
2 (f1 (y )dy + f2 (z )dz ) do
Suppose at the current iterate k, a feasible solution α = η · α.
(y (k) , z (k) ) is available. For every t ∈ T , let end while
λt−1 − λt yt λt R Output αk = α.
f1 (yt ) = e , f2 (zt ) = 2 (ezt − 1)2 ; End
τ τ
0 λt−1 − λt yt 0 λt R 2zt
f1 (yt ) = e , f2 (zt ) = 2 (2e − 2ezt ); We point out that in the LSP algorithm 1, α is first set as
τ τ
00 λt−1 − λt yt 00 λt R 2zt the maximal allowable step size αmax . In our experiments, we
f1 (yt ) = e , f2 (zt ) = 2 (4e − 2ezt ). choose αmax = 1. We have
τ τ
As pointed out in the previous subsection, the function f2 (zt ) Theorem 2. The line search algorithm 1 can always find a
is convex w.r.t. zt as ezt ≥ 21 under the setting for the step size α satisfying the inequality (8).
problem in this paper. Therefore, we can approximate f2 (zt )
in a neighborhood of ztk via its Taylor expansion as follows: Proof. See Appendix A-A.
Now we are ready to describe the following sequential
¯ k t,k k 0 k t,k f200 (ztk ) t,k 2 linear/quadratic programming (SLQP) approach for model (6)
f2 (zt + dz ) = f2 (zt ) + f2 (zt )dz + (dz ) .
2 as shown in Algorithm 2. The SLQP is a modified variant
Since the function f1 (yt ) is only convex when λt−1 ≥ λt , and of the classical sequential quadratic programming [27]. To
non-convex when λt−1 < λt , we suggest to approximate f1 (.) explore the conditions under which the solution generated by
(k) the SLQP method is a local minimum, we need to rewrite the
in some neighborhood of yt via the following function
( two box constraints (6e-6f) in the LCEO model as convex
k 0 k k,t f100(ytk) k,t 2
f
f¯1(yt +dy ) = 1 t
k k,t (y )+f (y
1 t y)d + 2 (dy ) If λt−1 ≥ λ t ; quadratic constraints. We can also eliminate the decision
f1 (ytk ) + f10 (ytk )dk,t
y Otherwise. variables {zt : t ∈ T } in the LCEO model via using the
 6

relation zt = yt+1 − yt . This leads to the optimization model Theorem 3. The SLQP algorithm 2 will generate a sequence
(9a). {(ytk , ztk )} converging to a stationary point (y ∗ , z ∗ ) of the
LCEO problem (6). Moreover, if (y ∗ , z ∗ ) satisfies the regularity
Algorithm 2 Sequential Linear/Quadratic Programming condition 1, then (y ∗ , z ∗ ) is a local minimum of LCEO
Input: Parameters: λt , τ, R, ȳ, y l , y u , z l , z u ,  > 0; model (6).
Output: Solution (ŷ, ẑ).
Proof. See Appendix A-B
Begin
Set k = 0; V. E XPERIMENTAL R ESULTS
Set ytk = ȳ, ∀1 ≤ t ≤ T + 1;
Set z k = 0, ∀1 ≤ t ≤ T ; In this section, we first describe the data set used in our
2 2 experiments, then run the simulation with 5-minute resolution
while k == 0 or y k −y k−1 + z k −z k−1 ≥  do
to validate the efficacy of our new model and algorithms under
Solve model (7) for (dky , dkz );
different scenarios. Finally, we report all the simulation results.
Use Algorithm 1 to find a suitable step size α(k) ;
All numerical experiments are conducted on a laptop with 2.2
Update (y k+1 = y k + α(k) dky , z k+1 = z k + α(k) dkz );
GHz six-core CPU and 16 GB memory. We use the 5-minute
k = k + 1;
electricity price in year 2020 from MISO real-time market
end while
[28]. We use the battery data from Berrueta’s paper [22]. Let
Output (x̂, ŷ) = (x(k) , y (k) ).
End xα,a = 0.083 + 0.917 · SOC; xα,c = 1 − 0.7 · SOC.
The OCVDF used in [22] is defined as follows:
 
X λt−1 − λt λt R 0 0 R·T (1 − xα,c ) · xα,a
min eyt + 2 (eyt+1 −yt − 1)2 (9a) veq = Uc − Ua + · ln
yt τ τ F xα,c · (1 − xα,a )
t∈T N
" #
s.t. y1 = yT +1 = ȳ (9b)
X k+1 2 · xα,c · k · (1 − xα,c )
+ Ak · (2 · xα,c − 1) − 1−k
l u
(yt − y )(yt − y ) ≤ 0, ∀t ∈ T ; (9c) k=0 (2 · xα,c − 1)
N
" #
l u
(yt+1 − yt − z )(yt+1 − yt − z ) ≤ 0, ∀t ∈ T . (9d) X k+1 2 · xα,a · k · (1 − xα,a )
− Ak · (2 · xα,a − 1) − 1−k
.
k=0 (2 · xα,a − 1)
Now let us consider the following Lagrangian function for the
above problem: For specific parameter values, we refer to the paper [22].
X λt−1 − λt λt R
L(yt , γt , ρt , νt ) = eyt + 2 (eyt+1 −yt − 1)2 A. VIAM VS LECO Simulation Results
τ τ
t∈T
In this subsection, we compare the performance of the
+ γ1 (y1 − ȳ) + γT +1 (yT +1 − ȳ)
X LCEO model, the VIAM model with a linear OCVDF (VIAM-
+ ρt (yt − y l )(yt − y u ) L) and the VIAM model with a nonlinear OCVDF (VIAM-
t∈T
X NL). To solve these models in reasonably large scale, we
+ νt (yt+1 − yt − z l )(yt+1 − yt − z u ). develop models and algorithms for the VIAM and LCEO by
t∈T using Julia [29] with IPOPT [26] as the solver for the sub-
problems in the SLQP approach. In our experiments, the scale
where (γt , ρt , νt ) are the Lagrangian multipliers for constraints
of the problem, measured by cardinality of T , is determined
(9b), (9c) and (9d), respectively. We introduce the following
by the number of days and price data resolution. For example,
regularity condition.
consider a scenario where the time span is 7 days and the price
Condition 1 (Regularity). We say the solution (y ∗ ) satisfies data resolution is 5-minutes, we have T = 7 ∗ 24 ∗ 12 = 2016.
the regularity condition, if To assess the effect from the fluctuation in electricity price
on the profit, we solve the daily profits through year 2020,
H = ∇2y L(y, γ, ρ, ν)  0,
and calculate the mean and the standard deviation of the daily
where H is a tri-diagonal matrix whose elements are defined profits from each model. Table I compares the performance of
as follows: three models (i.e., VIAM-L, VIAM-NL, and LCEO) in terms
of the profit based on the control strategy derived from the
2Re−2(yt−1 +yt ) obtained solution to the corresponding optimization model,
Ht,t = [λt−1 e3yt (2eyt − eyt−1 )
τ2 where the average profits and the standard deviations are
+ λt e2yt−1 +yt+1 (2eyt+1 − eyt )] computed for a normalized battery with the capacity of 1 MW
(λt−1 − λt ) eyt per day.
+ 2νt−1 + 2νt + 2ρt +
τ From Table I we can observe that on average, the same
2λt Reyt+1 −2yt (eyt − 2eyt+1 ) amount of profit can be made based on the optimal solutions
Ht,t+1 = Ht+1,t = − 2νt
τ2 of both the LCEO and the VIAM-L models, while a slightly
We have the following result. higher profit can be made based on the optimal solution to
the VIAM-NL model. This shows that the usage of the linear
 7

TABLE I: Solutions from VIAM and LECO Models lowest profits and plot the daily price dynamics for these 6
Days T Model AVG Profit SD of Profit days in Figure 5 and 6. As shown in Figure 5 and 6, a lower
VIAM-L 59.2439 66.6185 price volatility usually leads to a lower profit. As pointed out
1 288 VIAM-NL 59.3645 66.8112
LCEO 59.2439 66.6186
in Section III, such a conclusion can also be drawn from the
VIAM-L 59.7943 33.5843 LCEO model (6) directly.
7 2016 VIAM-NL 59.9172 33.6756
LCEO 59.7943 33.5844
VIAM-L 60.2586 26.3356
30 8640 VIAM-NL 60.3769 26.3337
LCEO 60.2532 26.2712
VIAM-L 60.2586 13.4307
90 25920 VIAM-NL 60.3823 13.4425
LCEO 60.2532 13.426
VIAM-L 60.2614 13.5676
180 51840 VIAM-NL 60.3881 13.565
LCEO 60.2532 13.5666
VIAM-L 59.5409 NA
366 105408 VIAM-NL 59.6659 NA
LCEO 59.5409 NA
Fig. 4: Daily Arbitrage Profits in 2020

approximation to the nonlinear OCVDF caused very few loss

in profit.
Figure 3 shows the average CPU time used to solve each
model versus the sizes of the solved instances in Julia. As we
can see from Figure 3, the VIAM-NL model always requires
the longest CPU time to solve, which indicates that the linear
approximation to the nonlinear OCVDF does help to reduce
substantially the difficulty of the underlying optimization
model. We also observe that in Figure 3, while the CPU
time used to solve each model increases as the problem size
increases, and the proposed SLQP approach for the LCEO Fig. 5: Daily Price dynamics in Most Profitable Time
enjoys the slowest growth rate of the CPU time. Particularly,
for the largest instance with over 300,000 decision variables,
the CPU time used by the SLQP for the LCEO is less than
1
3 of the CPU time used by IPOPT to solve the VIAM-L.
This indicates that the SLQP is more efficient for large scale
instances.

Fig. 6: Daily Price Dynamics in Least Profitable Time

Figure 7 shows the optimal BESS control strategy and

MISO electricity price on July 10, 2021. This strategy follows
a simple principal, i.e., to make more profits, the BESS should
discharge the battery at high prices and charge the battery at
low prices.
Fig. 3: Average CPU Time for VIAM and LECO Models
VI. C ONCLUSION
In this paper, we proposed a new BESS optimization
B. Impact of Price Dynamics on LECO Solutions model which naturally addresses the simultaneous charg-
In this subsection, we discuss the impact of price dynamics ing/discharging problem by using a single current decision
on the LCEO model solutions and profits. Figure 4 shows the variable. We also proposed to use some linear function to
daily profits of a normalized battery under MISO and the daily approximate the highly nonlinear OCVDF to simplify the
price dynamic in 2020. As shown in Figure 4, there is a high underlying BESS optimization model, which further leads to
volatility in the daily profits through the year. To assess the the new BLCQP model. To deal with the bilinear constraints
impact from the price dynamics on the daily profit, we select in the BLCQP model, we introduced a novel reformulation
three days with the highest profits and three days with the technique to recast it as another equivalent LCEO with only
 8

algorithm converges to a stationary point (y ∗ , z ∗ ) of model (6).

For notational simplicity, we convert constraints (6e-6f) to
equivalent convex constraints (yt − y l )(yt − y u ) ≤ 0 and
(zt − z l )(zt − z u ) ≤ 0. We note the constraints (6b-6f)
satisfy the linear independence constraint qualification (LICQ).
Because (y ∗ , z ∗ ) is a stationary point, it satisfies the following
KKT conditions.
λt−1 − λt yt
e − µt−1 + µt + ρt (2yt − y l − y u ) = 0, ∀t ∈ T ;
τ
2λt Req zt
(e − 1) ezt + µt + νt (2zt − z l − z u ) = 0, ∀t ∈ T ;
Fig. 7: SOC and Electricity Price on July 10, 2021 τ2
ρt yt − y l (yt − y u ) = 0, ρt ≥ 0, ∀t ∈ T ;

νt zt − z l (zt − z u ) = 0, νt ≥ 0, ∀t ∈ T ;


linear constraints. A new SLQP approach is proposed for Constraints (6b − 6f )
the LCEO model and its convergence is established. Our
preliminary experiments illustrate that the optimal solution The SLQP algorithm 2 generatesP (ŷ, ẑ), which minimizes the
from the new model provides a very good approximation to quadratic objective function t∈T f¯1 (yt ) + f¯2 (zt ), subject to
the optimal solution of the BESS optimization model with the the constraints (6b-6f). Therefore, (ŷ, ẑ) satisfies the following
original nonlinear OCVDF, and the proposed SLQP algorithm KKT conditions.
is competitive with the state-of-the-art optimization software 0
such as IPOPT. f¯1 (yt ) − µt−1 + µt + ρt (2yt − y l − y u ) = 0, ∀t ∈ T ;
0
f¯2 (zt ) + µt + νt (2zt − z l − z u ) = 0, ∀t ∈ T ;
A PPENDIX A ρt yt − y l (yt − y u ) = 0, ρt ≥ 0, ∀t ∈ T ;


P ROOFS OF THE T HEOREMS
νt zt − z l (zt − z u ) = 0, νt ≥ 0, ∀t ∈ T ;


A. Proof of Theorem 2
Constraints (6b − 6f )
PProof.0 kBecause (dk,t k,t
y , dz ) is a descent direction,
t,k 0 k t,k Because
t∈T f1 (yt )dy + f2 (zt )dz ≤ 0. By Taylor’s approxima-
tion, for sufficiently small α > 0, we have (
f10 (yˆt ) + f100 (yˆt )(yˆt − yˆt ) If λt−1 ≥ λt
X f¯10 (yˆt ) =
f1 (ytk + αdk,t k t,k
y ) + f2 (zt + αdz ) f10 (yˆt ) Otherwise.
t∈T
X X = f10 (yˆt )
f1 (ytk ) f2 (ztk ) f10 (ytk )dt,k f20 (ztk )dt,k


= + + α +
y z f¯20 (zˆt ) = f20 (zˆt ) + f200 (zˆt )(zˆt − zˆt ) = f20 (zˆt )
t∈T t∈T
!
X we observe the above KKT conditions coincide. Therefore,
+o α f10 (ytk )dt,k
y + f20 (ztk )dt,k
z (ŷ, ẑ) is a stationary point for model (6).
t∈T
X Xα Next, we prove that the stationary point is a local minimum.
f1 (ytk ) + f2 (ztk ) + f10 (ytk )dt,k 0 k t,k


= y + f2 (zt )dz According to textbook [27], the local minimality holds true if
2
t∈T
Xα
t∈T the second order sufficient conditions are satisfied. We note
f 0 (y k )dt,k + f20 (ztk )dt,k that Condition 1 implies the second order sufficient conditions,


+
2 1 t y z
this completes the proof.
t∈T
!
X
0 k t,k 0 k t,k
+o α f1 (yt )dy + f2 (zt )dz
t∈T
X α 0 k t,k R EFERENCES
f1 (ytk ) + f2 (ztk ) + f1 (yt )dy + f20 (ztk )dt,k


≤ z .
2 [1] R. Jiang, J. Wang, and Y. Guan, “Robust unit commitment with
t∈T
wind power and pumped storage hydro,” IEEE Transactions on Power
This completes the proof. Systems, vol. 27, no. 2, pp. 800–810, 2011.
[2] D. Mohler and D. Sowder, “Energy storage and the need for flexibility
on the grid,” in Renewable Energy Integration. Elsevier, 2017, pp.
B. Proof of Theorem 3 309–316.
[3] W. Cole, A. W. Frazier, and C. Augustine, “Cost projections for
Proof. To start, we observe that the line search procedure 1 utility-scale battery storage: 2021 update,” National Renewable Energy
can always find an α satisfying the relation (8). Moreover, Lab.(NREL), Golden, CO (United States), Tech. Rep., 2021.
[4] M. R. Sheibani, G. R. Yousefi, M. A. Latify, and S. H. Dolatabadi,
the termination of SLQP algorithm 2 is guaranteed, since “Energy storage system expansion planning in power systems: a review,”
the objective function value is monotonically decreasing and IET Renewable Power Generation, vol. 12, no. 11, pp. 1203–1221, 2018.
have finite lower bound (noting the box constraints for yt and [5] X. Luo, J. Wang, M. Dooner, and J. Clarke, “Overview of current
development in electrical energy storage technologies and the application
zt ). Next, we prove the SLQP algorithm converges to a local potential in power system operation,” Applied energy, vol. 137, pp. 511–
minimum point (y ∗ , z ∗ ) of model (6). We first prove the SLQP 536, 2015.
 9

[6] M. Stecca, L. R. Elizondo, T. B. Soeiro, P. Bauer, and P. Palensky,

“A comprehensive review of the integration of battery energy storage
systems into distribution networks,” IEEE Open Journal of the Industrial
Electronics Society, vol. 1, pp. 46–65, 2020.
[7] Y. Wang, Y. Dvorkin, R. Fernandez-Blanco, B. Xu, T. Qiu, and D. S.
Kirschen, “Look-ahead bidding strategy for energy storage,” IEEE
Transactions on Sustainable Energy, vol. 8, no. 3, pp. 1106–1117, 2017.
[8] S. Babaei, C. Zhao, and L. Fan, “A data-driven model of virtual power
plants in day-ahead unit commitment,” IEEE Transactions on Power
Systems, vol. 34, no. 6, pp. 5125–5135, 2019.
[9] S. Hajiaghasi, A. Salemnia, and M. Hamzeh, “Hybrid energy storage
system for microgrids applications: A review,” Journal of Energy Stor-
age, vol. 21, pp. 543–570, 2019.
[10] J. M. Arroyo, L. Baringo, A. Baringo, R. Bolaños, N. Alguacil, and N. G.
Cobos, “On the use of a convex model for bulk storage in MIP-based
power system operation and planning,” IEEE Transactions on Power
Systems, vol. 35, no. 6, pp. 4964–4967, 2020.
[11] Z. Li, Q. Guo, H. Sun, and J. Wang, “Sufficient conditions for exact re-
laxation of complementarity constraints for storage-concerned economic
dispatch,” IEEE Transactions on Power Systems, vol. 31, no. 2, pp.
1653–1654, 2015.
[12] S. Wang, G. Geng, and Q. Jiang, “Robust co-planning of energy storage
and transmission line with mixed integer recourse,” IEEE Transactions
on Power Systems, vol. 34, no. 6, pp. 4728–4738, 2019.
[13] Y. Wen, C. Guo, H. Pandžić, and D. S. Kirschen, “Enhanced security-
constrained unit commitment with emerging utility-scale energy stor-
age,” IEEE Transactions on power Systems, vol. 31, no. 1, pp. 652–662,
2015.
[14] A. J. Gonzalez-Castellanos, D. Pozo, and A. Bischi, “Non-ideal linear
operation model for li-ion batteries,” IEEE Transactions on Power
Systems, vol. 35, no. 1, pp. 672–682, 2019.
[15] H. Pandžić and V. Bobanac, “An accurate charging model of battery
energy storage,” IEEE Transactions on Power Systems, vol. 34, no. 2,
pp. 1416–1426, 2018.
[16] A. Sakti, K. G. Gallagher, N. Sepulveda, C. Uckun, C. Vergara, F. J.
de Sisternes, D. W. Dees, and A. Botterud, “Enhanced representations
of lithium-ion batteries in power systems models and their effect on the
valuation of energy arbitrage applications,” Journal of Power Sources,
vol. 342, pp. 279–291, 2017.
[17] P. Shrivastava, T. K. Soon, M. Y. I. B. Idris, and S. Mekhilef, “Overview
of model-based online state-of-charge estimation using Kalman filter
family for lithium-ion batteries,” Renewable and Sustainable Energy
Reviews, vol. 113, p. 109233, 2019.
[18] H. Hinz, “Comparison of lithium-ion battery models for simulating
storage systems in distributed power generation,” Inventions, vol. 4,
no. 3, p. 41, 2019.
[19] D. M. Rosewater, D. A. Copp, T. A. Nguyen, R. H. Byrne, and
S. Santoso, “Battery energy storage models for optimal control,” IEEE
Access, vol. 7, pp. 178 357–178 391, 2019.
[20] M. Shen and Q. Gao, “A review on battery management system from the
modeling efforts to its multiapplication and integration,” International
Journal of Energy Research, vol. 43, no. 10, pp. 5042–5075, 2019.
[21] J. Tan and Y. Zhang, “Coordinated control strategy of a battery energy
storage system to support a wind power plant providing multi-timescale
frequency ancillary services,” IEEE Transactions on Sustainable Energy,
vol. 8, no. 3, pp. 1140–1153, 2017.
[22] A. Berrueta, A. Urtasun, A. Ursúa, and P. Sanchis, “A comprehensive
model for lithium-ion batteries: From the physical principles to an
electrical model,” Energy, vol. 144, pp. 286–300, 2018.
[23] E. Zeidler, Nonlinear functional analysis and its applications: III:
variational methods and optimization. Springer Science & Business
Media, 2013.
[24] M. Grant and S. Boyd, “CVX: Matlab software for disciplined convex
programming, version 2.1,” 2014.
[25] D. G. Luenberger, Y. Ye et al., Linear and nonlinear programming.
Springer, 1984, vol. 2.
[26] A. Wächter and L. T. Biegler, “On the implementation of an interior-
point filter line-search algorithm for large-scale nonlinear programming,”
Mathematical programming, vol. 106, no. 1, pp. 25–57, 2006.
[27] J. Nocedal and S. Wright, Numerical optimization. Springer Science
& Business Media, 2006.
[28] MISO, “Midcontinent Independent System Operator (MISO).” [Online].
Available: https://www.misoenergy.org/
[29] J. Bezanson, A. Edelman, S. Karpinski, and V. B. Shah, “Julia: A fresh
approach to numerical computing,” SIAM Review, vol. 59, no. 1, pp.
65–98, 2017.