IEEE Transactions on Power Systems 1

Microgrid Optimal Energy Scheduling

Considering Neural Network based
Battery Degradation
Cunzhi Zhao, Student Member, IEEE and Xingpeng Li, Senior Member, IEEE

Abstract—Battery energy storage system (BESS) can effective- 𝑀𝑀𝑀𝑀𝑀𝑀 Maximum thermal limit of tie-line between main grid
𝑃𝑃𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺
ly mitigate the uncertainty of variable renewable generation. and microgrid.
Degradation is unpreventable and hard to model and predict for
𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅
𝑃𝑃𝐺𝐺𝐺𝐺 Ramping limit of diesel generator i.
batteries such as the most popular Lithium-ion battery (LiB). In 𝑃𝑃𝑆𝑆𝑆𝑆𝑀𝑀𝑀𝑀𝑀𝑀 Maximum charge/discharge power of BESS i.
this paper, we propose a data driven method to predict the bat- 𝑃𝑃𝑆𝑆𝑆𝑆𝑀𝑀𝑀𝑀𝑀𝑀 Minimum charge/discharge power of BESS i.
tery degradation per a given scheduled battery operational pro- 𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷
𝜂𝜂𝑆𝑆𝑆𝑆 Discharge efficiency of BESS i.
file. Particularly, a neural network based battery degradation 𝐶𝐶ℎ𝑎𝑎𝑎𝑎 Minimum capacity of generator i.
𝜂𝜂𝑆𝑆𝑆𝑆
(NNBD) model is proposed to quantify the battery degradation
with inputs of major battery degradation factors. When incorpo-
Variables:
rating the proposed NNBD model into microgrid day-ahead 𝑡𝑡
𝑈𝑈𝐵𝐵𝐵𝐵𝐵𝐵 Status of buying power from main grid in time interval t.
scheduling (MDS), we can establish a battery degradation based
MDS (BDMDS) model that can consider the equivalent battery
𝑡𝑡
𝑈𝑈𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 Status of selling power to main grid status in time t.
degradation cost precisely with the proposed cycle based battery 𝑡𝑡,𝑖𝑖 Charging status of energy storage system i in time inter-
𝑈𝑈𝐶𝐶ℎ𝑎𝑎𝑎𝑎
usage processing (CBUP) method for the NNBD model. Since the val t. It is 1 if charging status; otherwise 0.
proposed NNBD model is highly non-linear and non-convex, 𝑡𝑡,𝑖𝑖 Discharging status of energy storage system i in time
𝑈𝑈𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷
BDMDS would be very hard to solve. To address this issue, a interval t. It is 1 if discharging status; otherwise 0.
neural network and optimization decoupled heuristic (NNODH) 𝑡𝑡𝑡𝑡 Status of generator i in time interval t. It is 1 if on status;
𝑈𝑈𝐺𝐺
algorithm is proposed in this paper to effectively solve this neural otherwise 0.
network embedded optimization problem. Simulation results 𝑡𝑡𝑡𝑡 Startup indicator of Status of generator i in time interval
𝑉𝑉𝐺𝐺
demonstrate that the proposed NNODH algorithm is able to ob- t. It is 1 if unit i starts up; otherwise 0.
tain the optimal solution with lowest total cost including normal
𝑡𝑡
𝑃𝑃𝐺𝐺𝐺𝐺 Output of generator i in time interval t.
operation cost and battery degradation cost. 𝑡𝑡 Amount of power purchased from main grid power in
𝑃𝑃𝐵𝐵𝐵𝐵𝐵𝐵
Index Terms— Battery degradation, Battery energy storage time interval t.
system, Energy management system, Machine learning, Mi- 𝑡𝑡 Amount of power sold to main grid power in time inter-
𝑃𝑃𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆
crogrid day-ahead scheduling, Neural network, Optimization. val t.
𝑃𝑃𝐿𝐿𝑡𝑡 Demand in the microgrid in time interval t.
NOMENCLATURE
𝑡𝑡,𝑖𝑖
𝑃𝑃𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 Discharging power of energy storage system i at time t.
Sets:
𝑡𝑡,𝑖𝑖
𝑃𝑃𝐶𝐶ℎ𝑎𝑎𝑎𝑎 Charging power of energy storage system i at time t.
𝑆𝑆𝑇𝑇 Set of time intervals.
Set of controllable micro generators. Abbreviations:
𝑆𝑆𝐺𝐺
RES Renewable energy sources.
𝑆𝑆𝑆𝑆 Set of energy storage systems.
BESS Battery energy storage system.
𝑆𝑆𝑊𝑊𝑊𝑊 Set of wind turbines. LiB Li-ion battery.
𝑆𝑆𝑃𝑃𝑃𝑃 Set of PV systems. SOC State of charge.
SOH State of health.
Parameters: DOD Depth of discharge.
𝑐𝑐𝐺𝐺𝐺𝐺 Linear cost for controllable unit i. EV Electric Vehicle
𝑁𝑁𝑁𝑁
𝑐𝑐𝐺𝐺𝐺𝐺 No load cost for controllable unit i. MDS Microgrid day-ahead scheduling.
𝑆𝑆𝑆𝑆
𝑐𝑐𝐺𝐺𝐺𝐺 Start-up cost for controllable unit i. NNBD Neural network based battery degradation.
∆𝑇𝑇 Length of a single dispatch interval. BDMDS Battery degradation based MDS.
𝑅𝑅𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝 Percentage of the backup power to the total power. CBUP Cycle based battery usage processing.
𝑀𝑀𝑀𝑀𝑀𝑀
𝐸𝐸𝑆𝑆𝑆𝑆 Maximum energy capacity of ESS i. NNODH Neural network and optimization decoupled heuristic.
𝑚𝑚𝑚𝑚𝑚𝑚 Minimum energy capacity of ESS i. CMDS Conserved MDS.
𝐸𝐸𝑆𝑆𝑆𝑆
𝑡𝑡
𝑐𝑐𝐵𝐵𝐵𝐵𝐵𝐵 Wholesale electricity purchase price in time interval t.
𝑡𝑡
𝑐𝑐𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 Wholesale electricity sell price in time interval t.
I. INTRODUCTION

R
𝑀𝑀𝑀𝑀𝑀𝑀
𝑃𝑃𝐺𝐺𝐺𝐺 Maximum capacity of generator i.
𝑀𝑀𝑀𝑀𝑀𝑀
𝑃𝑃𝐺𝐺𝐺𝐺 Minimum capacity of generator i. enewable energy sources (RES) will play an important
role in future microgrid due to the 3-D (decentralization,
decarbonation, and digitalization) trend. The proportion of
Cunzhi Zhao and Xingpeng Li are with the Department of Electrical and RES in power generation is growing significantly. However,
Computer Engineering, University of Houston, Houston, TX, 77204, USA (e- the system stability is substantially weakened by the stochastic
mail: [email protected]; [email protected]).
IEEE Transactions on Power Systems 2

and intermittent generation of high penetration RES. Battery battery degradation is high, the developed model is hard to
energy storage system (BESS) is an effective flexible solution incorporate with the various operation of BESS due to the
for addressing the uncertainty of variable RES induced system limited battery degradation data. A temperature based battery
[1]. Thus, much more BESSs will be available in bulk power degradation model is proposed in [26] for electric vehicle
systems and small-scale microgrids in near future. (EV). However, the model is only reflected by the internal
The characteristic of rechargeable chemical battery makes it temperature of EV’s battery pack and may not suitable for the
degrade during cycling. This degradation can be accelerated BESS. Saldana’s paper [27] proposed a battery degradation
by extreme fast charging or discharging cycles, extreme low matrix reference for EV. However, it is impractical to consider
or high ambient temperature and over charging or over dis- this model in the microgrid day-ahead scheduling (MDS)
charging. However, the internal states of battery remain diffi- problem due to the computational complexity in the optimiza-
cult to estimate while the battery is taking a more important tion problem. The electric vehicle has been researched and
role as BESS in power energy systems [2]. Therefore, the bat- conducted in the V2G system in [28]-[30] due to the similarity
tery degradation is quite hard to predict. The working princi- between EV and BESS. However, an important EV battery
ple of a BESS is similar to a voltage source in series with im- degradation factor, the charge/discharge rate, is neglected in
pedances. However, the operating conditions and environ- those papers which may lead to an inaccurate battery degrada-
ments are various for BESS in microgrids, as well as in bulk tion prediction. A data driven degradation model is presented
power systems. Thus, the BESS model cannot be simply treat- in [31]. A quadratic equation is formed based on the collected
ed as a voltage-source based model [3]. Li-ion battery (LiB) data; however, the data are collected only under different pro-
has been widely used as energy storage due to its high energy files of DOD and SOC which omits other degradation related
density and low memory effect nature. LiB degrades mainly factors. References [32]-[34] present some advanced method-
because of the loss of Li-ions, the loss of electrolyte and the ology to predict the lifetime of the battery cell and then the
increase of internal resistance [4]. During battery cycling, the degradation can be averaged for each cycle. However, those
influential factors that cause the degradation include the bat- methodologies are specially designed for the battery aging
tery operating ambient temperature, charging/discharging rate, tests in which the battery’s cycle is set as a fixed charge or
state of charge (SOC), state of health (SOH), and depth of discharge rate. In other words, the battery degradation or re-
discharge (DOD) [5]-[6]. maining cycle prediction is based on the fixed charging or
Previous studies have developed some energy management discharging cycles in those methodologies. Therefore, they
strategies for BESS-integrated microgrids. In [7]-[14], it has cannot reflect the degradation prediction for usage-based
been proved that the BESS can be seamlessly integrated into BESS due to the battery’s dynamic schedule such as various
microgrid especially for those microgrids with high penetra- charge or discharge rates at different time intervals in power
tion renewable energy sources. However, battery degradation system/microgrid applications.
is not considered in those energy management strategies. In summary, there are mainly three gaps for all the afore-
Some simple models of battery degradation are proposed in mentioned battery degradation models, which are addressed in
the literature. In [15], DOD is used to calculate the remaining this paper:
useful cycles. The battery degradation is estimated based on 1) Existing models do not consider all major critical deg-
the remaining useful battery cycle and the actual capacity. The radation factors. They focus on only a couple of varia-
only variable considered in [15] is the DOD and they assume bles and ignore other critical factors, which limits the
the degradation process to be linear throughout the battery life
accuracy of their battery degradation models.
which is not reasonable. Similar to [15], [16]-[19] proposed
some DOD based models to estimate the battery degradation 2) BESS operations are often very dynamic; for instance,
for each cycle which omits other important degradation fac- they may have very different charge/discharge rates
tors. Pascali in [20] adopts the Butler-Volmer equation for and SOC levels at different time intervals. This is not
battery degradation model to illustrate the diffusion of the respected by existing methods that are unable to accu-
solvent reactants. However, the degradation data are based on rately consider such dynamics in usage-based battery
the experiment for different discharge currents and SOC val- degradation prediction.
ues, which are not sufficient as they omit other important fac-
3) Some existing degradation models are not positioned to
tors contributing to battery degradation. As mentioned in [21]-
[24], the linear assumption of battery degradation may simpli- be efficiently incorporated into MDS.
fy the problem and reduce the computational difficulty, but the To address the above-mentioned gaps, a fully connected
degradation value may not be accurately predicted. In sum- neural network (NN) is proposed to train a battery degradation
mary, these popular heuristic battery degradation models can model, and a cycle based battery usage processing (CBUP)
be represented as two battery degradation models: (i) linear method is developed for the BESS scheduling to accurately
degradation model that applies a linear degradation cost based predict the battery degradation with the proposed NN model.
on the power output or the energy usage which may lead to a The input of the NN model is a vector of five features includ-
large error on battery degradation quantification, and (ii) DOD ing ambient temperature, charging/discharging rate, SOC,
and/or SOC based model that considers DOD and/or SOC as DOD and SOH. This neural network based battery degradation
the input which omits the other important degradation factors (NNBD) model contains non-linear activation functions in the
such as charge/discharge rate and ambient temperature. hidden layers, which makes it complex. When incorporating
A data driven, comprehensive battery degradation model is the proposed NNBD model into MDS, we can establish a bat-
developed in [25]. Although the accuracy of the prediction for
IEEE Transactions on Power Systems 3

tery degradation based MDS (BDMDS) model that can con- agement strategies with the BESS degradation model. Section
sider the equivalent battery degradation cost. V discusses the microgrid testbed, simulation results and sen-
The BESS scheduling in MDS does not always operate in a sitivity analysis. Section VI concludes the paper.
fixed charging or discharging cycle. Therefore, the CBUP
method is designed to fill the research gap between the fixed- II. TRADITIONAL MICROGRID DAY-AHEAD SCHEDULING
cycle based NNBD model and the dynamic BESS operation in A traditional MDS model is established as a basic model to
MDS. However, such a complex neural network embedded gauge the proposed BDMDS model. This traditional MDS
optimization problem would be hard to solve directly. To ad- model consists of (1)-(15) as described below and it does not
dress this issue, a neural network and optimization decoupled consider battery degradation.
heuristic (NNODH) algorithm is proposed in this paper to The objective of this traditional MDS model is to minimize
effectively solve the BDMDS problem. The proposed the total cost of the microgrid operations as illustrated in (1).
NNODH algorithm iteratively solves the transformed BDMDS The power balance equation involving controllable generators,
problem that is decoupled to the battery degradation calcula- renewable energy sources, power exchange with the main
tion and MDS optimization problems. BESS operation con- grid, BESS output and the load is shown in (2). Constraint (3)
straints with tighter bounds will be generated in each iteration enforces the power limits of the controllable units such as die-
to limit the usage of BESS which can reduce the battery deg- sel generators. The ramping up and down limits are enforced
radation and the relevant cost in the next iteration. However, by (4) and (5). Equation (6) ensures the status of power ex-
the microgrid’s operation cost will increase if the BESS usage change between microgrid and main grid to be either purchas-
is limited. The goal of the proposed NNODH algorithm is to ing or selling or stay idle. The thermal limit of the tie-line is
find the lowest value for the sum of battery degradation cost enforced by constraints (7)-(8). Equation (9) restricts the
and microgrid operation cost. It can also record the total cost BESS to be either in charging mode or in discharging mode or
for each iteration and locate the vertex point which is also the stay idle. Constraints (10)-(11) limit the charging/discharging
optimal solution for the BDMDS problem. Three benchmark power of BESS. As shown in (12), the SOC level of BESS can
models are also developed to test and compare the perfor- be calculated based on current energy stored in BESS. Equa-
mance of the proposed NNODH algorithm. The main contri- tion (13) calculates the energy stored in the BESS for each
butions of this paper are summarized as follows: time interval. The ending SOC level of BESS is forced to be
• A set of battery cycle generators is designed to simulate equal to the initial SOC value (14). Constraint (15) ensures the
battery degradation under different battery operational microgrid to have sufficient backup power to address outage
profiles. The key features (ambient temperature, events.
charge/discharge rate, SOC level, DOD and updated bat- Objective:
tery energy capacity) that affect battery degradation are 𝑡𝑡
𝑓𝑓 𝑀𝑀𝑀𝑀 = � �(𝑃𝑃𝐺𝐺𝐺𝐺 𝑁𝑁𝑁𝑁
𝑐𝑐𝐺𝐺𝐺𝐺 + 𝑈𝑈𝐺𝐺𝐺𝐺 𝑐𝑐𝐺𝐺𝐺𝐺 𝑆𝑆𝑆𝑆
+ 𝑉𝑉𝐺𝐺𝐺𝐺 𝑐𝑐𝐺𝐺𝐺𝐺 )
collected for each cycle. (1)
𝑡𝑡 𝑡𝑡 𝑡𝑡 𝑡𝑡
• A neural network based battery degradation model with + 𝑃𝑃𝐵𝐵𝐵𝐵𝐵𝐵 𝑐𝑐𝐵𝐵𝐵𝐵𝐵𝐵 − 𝑃𝑃𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 𝑐𝑐𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 , ∀𝑖𝑖, 𝑡𝑡
the above five input features is proposed in this paper and Constraints are as follows:
it is able to accurately predict the degradation respect to 𝑡𝑡
𝑃𝑃𝐵𝐵𝐵𝐵𝐵𝐵 +� 𝑡𝑡
𝑃𝑃𝐺𝐺𝐺𝐺 +� 𝑡𝑡
𝑃𝑃𝑊𝑊𝑊𝑊𝑊𝑊 +� 𝑡𝑡
𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃
the current maximum battery energy capacity. 𝑖𝑖∈𝑆𝑆𝐺𝐺 𝑖𝑖∈𝑆𝑆𝑊𝑊𝑊𝑊 𝑖𝑖∈𝑆𝑆𝑃𝑃𝑃𝑃
• A cycle based battery usage processing method is pro- + � 𝑡𝑡,𝑖𝑖
𝑃𝑃𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷
posed to process the BESS profile to correctly incorpo- 𝑖𝑖∈𝑆𝑆𝑆𝑆 (2)
rate the proposed NNBD model into MDS, addressing 𝑡𝑡
= 𝑃𝑃𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 +� 𝑃𝑃𝐿𝐿𝐿𝐿𝑡𝑡
the inconsistence between the fixed-cycle based NNBD 𝑖𝑖∈𝑆𝑆𝐿𝐿
model and BESS scheduling. +� 𝑡𝑡,𝑖𝑖
𝑃𝑃𝐶𝐶ℎ𝑎𝑎𝑎𝑎
• A BDMDS model is proposed to capture the effect of 𝑖𝑖∈𝑆𝑆𝑆𝑆
battery degradation by incorporating the proposed NNBD 𝑀𝑀𝑀𝑀𝑀𝑀 𝑡𝑡
𝑃𝑃𝐺𝐺𝐺𝐺 ≤ 𝑃𝑃𝐺𝐺𝐺𝐺 ≤ 𝑃𝑃𝐺𝐺𝐺𝐺 𝑀𝑀𝑀𝑀𝑀𝑀
, ∀𝑖𝑖, 𝑡𝑡 (3)
model into microgrid energy management. 𝑡𝑡+1 𝑡𝑡
𝑃𝑃𝐺𝐺𝐺𝐺 − 𝑃𝑃𝐺𝐺𝐺𝐺 ≤ ∆𝑇𝑇 ∙ 𝑃𝑃𝐺𝐺𝐺𝐺
𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅
, ∀𝑖𝑖, 𝑡𝑡 (4)
• An NNODH algorithm is proposed to efficiently solve 𝑡𝑡 𝑡𝑡+1 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 (5)
𝑃𝑃𝐺𝐺𝐺𝐺 − 𝑃𝑃𝐺𝐺𝐺𝐺 ≤ ∆𝑇𝑇 ∙ 𝑃𝑃𝐺𝐺𝐺𝐺 , ∀𝑖𝑖, 𝑡𝑡
the battery degradation based MDS model that is hard to
solve directly. Four battery usage-limiting MDS models,
𝑡𝑡
𝑈𝑈𝐵𝐵𝐵𝐵𝐵𝐵 𝑡𝑡
+ 𝑈𝑈𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 ≤ 1, ∀𝑡𝑡 (6)
referred to as conserved MDS (CMDS), are developed 0 ≤ 𝑃𝑃𝐵𝐵𝐵𝐵𝐵𝐵 𝑡𝑡
≤ 𝑈𝑈𝐵𝐵𝐵𝐵𝐵𝐵 𝑡𝑡 𝑀𝑀𝑀𝑀𝑀𝑀
∙ 𝑃𝑃𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺 , ∀𝑡𝑡 (7)
and used by the NNODH algorithm. The optimal sched- 𝑡𝑡 𝑡𝑡
0 ≤ 𝑃𝑃𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 ≤ 𝑈𝑈𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 ∙ 𝑃𝑃𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺 , ∀𝑡𝑡 𝑀𝑀𝑀𝑀𝑀𝑀 (8)
uling obtained with NNODH leads to the lowest total 𝑡𝑡,𝑖𝑖
𝑈𝑈𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 𝑡𝑡,𝑖𝑖
+ 𝑈𝑈𝐶𝐶ℎ𝑎𝑎𝑎𝑎 ≤ 1, ∀𝑖𝑖, 𝑡𝑡 (9)
cost including the battery degradation cost and microgrid
operation cost.
𝑡𝑡,𝑖𝑖 𝑀𝑀𝑀𝑀𝑀𝑀 𝑡𝑡,𝑖𝑖
𝑈𝑈𝐶𝐶ℎ𝑎𝑎𝑎𝑎 ∙ 𝑃𝑃𝑆𝑆𝑆𝑆 ≤ 𝑃𝑃𝐶𝐶ℎ𝑎𝑎𝑎𝑎 ≤ 𝑈𝑈𝐶𝐶ℎ𝑎𝑎𝑎𝑎 𝑡𝑡,𝑖𝑖
∙ 𝑃𝑃𝑆𝑆𝑆𝑆𝑀𝑀𝑀𝑀𝑀𝑀 , ∀𝑖𝑖, 𝑡𝑡 (10)
• Validation of the performance for the proposed NNODH 𝑡𝑡,𝑖𝑖 𝑡𝑡,𝑖𝑖 𝑡𝑡,𝑖𝑖
𝑈𝑈𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 ∙ 𝑃𝑃𝑆𝑆𝑆𝑆𝑀𝑀𝑀𝑀𝑀𝑀 ≤ 𝑃𝑃𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 ≤ 𝑈𝑈𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 ∙ 𝑃𝑃𝑆𝑆𝑆𝑆𝑀𝑀𝑀𝑀𝑀𝑀 , ∀𝑖𝑖, 𝑡𝑡 (11)
algorithm is conducted. Case studies prove that by limit- 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑡𝑡 = 𝐸𝐸𝑆𝑆𝑆𝑆 𝑡𝑡 𝑀𝑀𝑀𝑀𝑀𝑀
⁄𝐸𝐸𝑆𝑆𝑆𝑆 , ∀𝑖𝑖, 𝑡𝑡 (12)
ing the battery operation that leads to lower degradation, 𝑡𝑡 𝑡𝑡−1 𝑡𝑡−1,𝑖𝑖 𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 𝑡𝑡−1,𝑖𝑖 𝐶𝐶ℎ𝑎𝑎𝑎𝑎
the total cost can be reduced significantly. 𝐸𝐸𝑆𝑆𝑆𝑆 − 𝐸𝐸𝑆𝑆𝑆𝑆 + ∆𝑇𝑇 ∙ �𝑃𝑃𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 �𝜂𝜂𝑆𝑆𝑆𝑆 − 𝑃𝑃𝐶𝐶ℎ𝑎𝑎𝑎𝑎 𝜂𝜂𝑆𝑆𝑆𝑆 � (13)
The rest of the paper is organized as follows. Section II pre- = 0, ∀𝑖𝑖, 𝑡𝑡
sents a traditional MDS model. Section III presents the neural
𝑡𝑡=24
𝐸𝐸𝑆𝑆𝑆𝑆 𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼𝐼
= 𝐸𝐸𝑆𝑆𝑆𝑆 , ∀𝑖𝑖 (14)
network structure of the proposed battery degradation model
and Section IV presents the proposed microgrid energy man-
IEEE Transactions on Power Systems 4

𝑀𝑀𝑀𝑀𝑀𝑀
𝑃𝑃𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺 𝑡𝑡
− 𝑃𝑃𝐵𝐵𝐵𝐵𝐵𝐵 𝑡𝑡
+ 𝑃𝑃𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 𝑀𝑀𝑀𝑀𝑀𝑀
+ � (𝑃𝑃𝐺𝐺𝐺𝐺 𝑡𝑡
− 𝑃𝑃𝐺𝐺𝐺𝐺 ) referred as C rate. As shown in (16), C rate measures the
speed at which a battery is fully charged or discharged.
𝐺𝐺 ∈ 𝑆𝑆𝐺𝐺
(15) 𝐼𝐼𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵
𝐶𝐶 𝑟𝑟𝑟𝑟𝑟𝑟𝑟𝑟 = (16)
𝐸𝐸0
≥ 𝑅𝑅𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝 � � 𝑃𝑃𝐿𝐿𝐿𝐿𝑡𝑡 � , ∀𝑖𝑖, 𝑡𝑡
where 𝐼𝐼𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 represents the charging/discharging current of
𝑖𝑖∈ 𝑆𝑆𝐿𝐿
the BESS and 𝐸𝐸0 denotes the rated capacity of BESS. For ex-
III. NEURAL NETWORK BASED BATTERY DEGRADATION ample, 2C means the battery will discharge the full capacity in
MODELING 0.5 hour. This work conducted 945 different battery aging
tests with different values of degradation factors.
The traditional microgrid day-ahead scheduling model de-
Table I Numbers of battery aging tests under various SOC and DOD levels
termines the optimal operational profiles for BESS, controlla-
ble generators and tie-line exchange power. However, the Initial SOC
BESS in traditional MDS model is considered to be ideal DOD 100% 80% 60% 50% 40% 20%
without any degradation and the equivalent battery degrada- 20% 4 5 17 23 22 23
tion cost is zero. This may substantially accelerate the aging 30% 20 34 36 32 37 /
and replacement of expensive BESS which may lead to eco- 40% 36 41 41 40 44 /
nomic loss in the long run. Besides, the BESS degradation 50% 38 41 36 42 / /
needs to be accurately quantified for various battery daily op- 60% 37 37 37 / / /
70% 41 36 / / / /
erational profiles to obtain the truly optimal scheduling solu-
80% 39 35 / / / /
tions. Thus, a deep learning method, particularly a deep neural
90% 35 / / / / /
network, is proposed in this section to accurately predict the 100% 36 / / / / /
BESS degradation.
B. Data Pre-processing
A. Input Data
The battery capacity level or SOH in percent at the end of
The deep neural network is applied to learn and predict the
each cycle is recorded from each simulation of battery aging
battery degradation with several critical features. A battery
test. Battery degradation per cycle is defined as the difference
model implemented in the MATLAB Simulink [35] is used to
between the initial SOH and ending SOH for a given cycle.
conduct the battery aging tests. The aging test model is mod-
However, for some battery aging tests with low C rates and
elled with the battery model and cycling generator in Sim-
small DODs, the degradation for some cycles may be too less
ulink. The SOC variable can be adjusted in the battery setting.
to measure and could even be zero. Also, some of the degrada-
The battery model of Simulink contains several default types
tion data contain outlier points which may require some fur-
of batteries and the battery parameters can also be adjusted
ther process for a better training result. The battery degrada-
with the manufacture’s datasheet if we want to introduce a
tion data are processed with different methods: (i) raw, (ii)
new type of battery. It can also simulate the ambient tempera-
smoothed method, and (iii) regressed method, as shown in Fig.
ture effects and aging effects of battery. A dynamic internal
1. “Raw” represents the original data without any pre-
resistance that is highly related to the battery degradation is
processing and the associated training results can be used as a
also simulated within the battery model. This battery model
benchmark to gauge the effectiveness of the other two pro-
can simulate various types and configurations of batteries,
posed data pre-processing methods. The smoothed method
various conditions, and operating profiles. Built upon this bat-
filters out the outliers of the raw data. Regressed method ap-
tery model, a battery cycle generator is designed to simulate
plies linear regression on the smoothed data. All three groups
the charging and discharging cycles under preset charg-
of input data are normalized to increase the training efficiency.
ing/discharging rates. For each battery aging test, each cycle is
The normalization used in [37] is applied in the data pre-
simulated to discharge from a certain SOC to a certain lower
processing and shown in (17):
SOC and then charge back to the starting SOC. The data col- 𝑥𝑥 −𝐸𝐸(𝑥𝑥𝑘𝑘 )
lected from the battery aging tests include the ambient temper- �𝑘𝑘 = 𝑘𝑘
𝑥𝑥 (17)
�𝑉𝑉𝑉𝑉𝑉𝑉(𝑥𝑥𝑘𝑘 )
ature, charging/discharging rate, SOC, DOD and reduced en- where 𝐸𝐸(𝑥𝑥𝑘𝑘 ) represents the expectation of 𝑥𝑥𝑘𝑘 and 𝑉𝑉𝑉𝑉𝑉𝑉(𝑥𝑥𝑘𝑘 )
ergy capacity level. The battery’s energy capacity level for represents the variance of the 𝑥𝑥𝑘𝑘 . The input data will be split
each cycle is also used to calculate each cycle’s battery degra- into two parts, 80% as the training dataset and 20% as the val-
dation value. Table I lists the numbers of battery aging tests idation dataset. At last, three groups of data will be fed into
that are simulated under different SOC and DOD. Each battery the NN separately to evaluate their performance.
aging test represents a group of simulated battery profile under Data preprocessing is conducted with the smoothed method
initial SOC and fixed DOD until the battery capacity degrades to reduce the random variation representing random measure-
to 80% of the maximum rated capacity. This indicates each ment errors, which makes it harder to train a neural network
battery aging test may contain different numbers of cycles. It model. Also, the MDS problem in this paper means that we
is common to consider the 80% of the rated capacity as the focus on the degradation cost for a relatively long period
end of battery’s life in battery aging test [36]. For instance, the which is typically 24 hours. Comparing with the unsmoothed
“4” in Table I means there are four battery aging tests that are data and the smoothed data, we found for the same cycle, the
simulated at an initial SOC level of 100% with a DOD level of total battery degradation is almost the same for a 24 hour time
10%. Moreover, each of these SOC&DOD combinations will period. However, the training accuracy of the NNBD model
be simulated under different ambient temperatures and charg- can be increased by smoothing out the input data while the
ing/discharging rates. The charging/discharging rate is also
IEEE Transactions on Power Systems 5

prediction error does not increase with the smoothed data. The
reason is that the unsmoothed data is quite unstable in some
short period as shown in the Fig. 1, which increases the diffi-
culty of machine learning model training and decrease the
training accuracy. However, if we look at the degradation for a
long time period, the smoothed data and the unsmoothed data
have a similar degradation value as shown in Table II. The
cumulative degradation numbers in Table II are calculated
based on the battery aging test with 60% SOC, 20% DOD,
0.75 C rate and 32℃ ambient temperature. There are five cy- Fig. 2 Structure of the proposed neural network model for battery degradation.
cle intervals analyzed in Table II; from the results, we can Mini batch gradient descent strategy is applied to train the
observe that the total degradation value for the certain interval neural network. Different batch sizes are tested to achieve the
between the raw data and the pre-processed data are very simi- best training results. As defined in (18), mean squared error
lar. Also, when the length of the interval increases, the differ- (MSE) represents the average of the square of the difference
ence between the raw and the pre-processed data decreases. between the actual and predicted values over all training data
This ensures that the data pre-processing method improves the points. MSE is used as the criterion, a loss function, to train
training accuracy without affecting the cumulative degradation the neural network and measure the neural network quality.
value. 𝑛𝑛
0.0034 1
RAW Regressed Smoothed 𝑀𝑀𝑀𝑀𝑀𝑀 = �(𝑦𝑦𝑖𝑖 − 𝑦𝑦�)
𝚤𝚤
2 (18)
𝑛𝑛
Degradation Percentage %

0.0032 𝑖𝑖=1

0.003 D. Battery Degradation Calculation

From the BESS operation profile, the SOC level is required
0.0028
as the input of the proposed NNBD model. The absolute value
0.0026 of the difference in the SOC levels between time intervals 𝑡𝑡
and 𝑡𝑡 − 1 will be the DOD level as shown in (19). C rate is
0.0024
calculated by (20). It is assumed that the battery SOH level is
0.0022 available prior to the microgrid day ahead scheduling. The
input vector can be formed as shown in (21) and then fed into
1

1001
1501
2001
2501
3001
3501
4001
4501
5001
5501
6001
6501
7001
7501
8001
501

Cycle Number the trained NNBD model to obtain the total battery degrada-
Fig. 1. Battery degradation data under different data processing methods. tion over the MDS time horizon in (22).
𝐷𝐷𝐷𝐷𝐷𝐷𝑡𝑡 = |𝑆𝑆𝑆𝑆𝑆𝑆𝑡𝑡 − 𝑆𝑆𝑆𝑆𝑆𝑆𝑡𝑡−1 | (19)
Table II The cumulative degradation comparison over raw data and pre- 𝐶𝐶𝑡𝑡𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 = 𝐷𝐷𝐷𝐷𝐷𝐷𝑡𝑡 /∆𝑇𝑇 (20)
processed data
Cycle Raw data Pre-processed
𝑥𝑥𝑡𝑡 = (𝑇𝑇, 𝐶𝐶, 𝑆𝑆𝑆𝑆𝑆𝑆, 𝐷𝐷𝐷𝐷𝐷𝐷, 𝑆𝑆𝑆𝑆𝑆𝑆) (21)
Difference 𝐵𝐵𝐵𝐵 = ∑𝑡𝑡∈𝑆𝑆𝑇𝑇 𝑓𝑓 𝑁𝑁𝑁𝑁 (𝑥𝑥𝑡𝑡 )𝑆𝑆𝑆𝑆𝑆𝑆 (22)
Range degradation degradation
0-500 1.213% 1.231% 1.5%
0-1000 2.454% 2.483% 1.18%
E. Cycle Based Battery Usage Processing Method
0-1500 3.711% 3.752% 1.10% The equation (22) for battery degradation calculation men-
0-2000 5.007% 5.048% 0.82% tioned in the previous section covers all the time periods for
0-2500 6.347% 6.362% 0.24% the microgrid day ahead scheduling. However, the BESS is
not scheduled in fixed cycles with constant charge/discharge
C. The Proposed NNBD Model rates per MDS, leading to inconsistency with the fixed-cycle
A fully connected neural network is constructed to model based data that are used for training NNBD model which is the
the battery degradation. Five aging factors (ambient tempera- gap mentioned in the introduction section. Furthermore, the
ture, C rate, SOC, DOD and SOH) form a five-element input idle status is also considered in (22) but it is out sampled for
vector for the neural network. Each input vector corresponds the NNBD model. In other words, equation (22) considers
to a single output value which is the amount of battery degra- each time interval as a BESS cycle and feeds the associated
dation in percentage respect to the SOH level for the same data directly into the NNBD model, which may not accurately
cycle. A dynamic learning rate scheme is used in the training predict the BESS degradation. Therefore, the proposed cycle
process to improve the training result. The learning rate will based battery usage processing method is developed to address
decrease automatically after a certain number of training the inconsistency and accurately apply the proposed NNBD
epochs. The structure of the trained neural network is shown model in the MDS problem.
in Fig. 2 plotted by NN-SVG [38]. It has an input layer with 5 In the CBUP method, instead of considering each time in-
neurons, first hidden layer with 20 neurons, second hidden terval as a cycle, the operating time intervals of BESS sched-
layer with 10 neurons and an output layer with 1 neuron. The uling are combined and averaged into different cycles. For any
activation function for the hidden layers is “relu” and “linear” continuous time intervals, if the operation status (charging or
for the output layer. discharging) does not change, they will be aggregated as a
single charging or discharging cycle. For the aggregated cycle,
the charging or discharging power will be the average of the
aggregated time periods. The SOC will be the initial SOC val-
IEEE Transactions on Power Systems 6

ue of the first time period in the aggregated time periods and Start
the DOD will be the absolute value of the SOC difference
between the start time period and the end time period respec- Step A: Solve Microgrid Day-Ahead Scheduling
tively. Thus, equation (22) is replaced by (23) while 𝑥𝑥𝑐𝑐 repre-
Step B: Obtain the BESS Schedule and
sents the input vector for the aggregated cycles and 𝐴𝐴𝐴𝐴 repre- Step E: Update Process it with the CBUP Method
sents the set of feasible aggregated cycles. BESS Operation
Constraints Step C: Estimate Battery Degradation with the
𝐵𝐵𝐵𝐵 = � 𝑓𝑓 𝑁𝑁𝑁𝑁 (𝑥𝑥𝑐𝑐 ) 𝑆𝑆𝑆𝑆𝑆𝑆 (23)
NNBD model and Calculate the Associated
𝑐𝑐∈𝐴𝐴𝐴𝐴
Equivalent Degradation Cost
F. Incorporating NNBD into Microgrid Scheduling
When considering battery degradation in microgrid day- No
Step D:
ahead scheduling, the objective function needs to be updated Stopping Criteria Met?
by including the associated equivalent battery degradation YES

cost. The updated objective function is shown in (24), End

𝑓𝑓 = 𝑓𝑓 𝑀𝑀𝑀𝑀 + 𝑓𝑓 𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 (24) Fig. 3. Flowchart of the proposed NNODH algorithm.
where 𝑓𝑓 is defined in equation (1); and 𝑓𝑓 𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 denotes the
𝑀𝑀𝑀𝑀
A. Microgrid Scheduling
battery degradation cost that can be estimated by the proposed The traditional microgrid day-ahead scheduling that intro-
NNBD model. 𝑓𝑓 𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 can be calculated as follows, duced in Section II is conducted to obtain an initial solution
𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶 𝑆𝑆𝑆𝑆
𝑓𝑓 𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 =
𝑐𝑐𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 −𝑐𝑐𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵
𝐵𝐵𝐵𝐵 (25) which does not consider the battery degradation. It is solved
1−𝑆𝑆𝑆𝑆𝑆𝑆𝐸𝐸𝐸𝐸𝐸𝐸 only in the first iteration. When considering battery usage lim-
𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶
where 𝑐𝑐𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 represents the capital investment cost of BESS; iting constraints that will be presented in Section IV.D, Step A
𝑆𝑆𝑆𝑆
𝑐𝑐𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 denotes the salvage value at the end of life; 𝑆𝑆𝑆𝑆𝑆𝑆𝐸𝐸𝐸𝐸𝐸𝐸 will solve the CMDS instead of the traditional MDS. To sum
represents state of health value that the BESS is considered as up, traditional MDS is solved in the first iteration while
the end of life; BD represents the percentage battery degrada- CMDS is solved in all subsequent iterations.
tion calculated by (23). B. Operation Limits of BESS
Thus, the proposed battery degradation based MDS model
The goal of Step E is to generate new constraints that can
can be represented by (2)-(15), (19)-(21), and (23)-(25).
limit or change the battery output to reduce the total battery
IV. THE PROPOSED NNODH FOR MICROGRID SCHEDULING degradation. The generated constraints will further tighten the
range that BESS can operate after each iteration. Three types
The proposed battery degradation based microgrid day- of extra constraints are proposed to limit battery output. Con-
ahead scheduling model, which is presented in the above sec- straint (26) named as battery consumption limit (BCL) is to
tion, would be very hard to solve directly since the proposed reduce the sum of battery output power over 24 hours by forc-
NNBD model that is highly non-linear and non-convex is now ing it to be less than the previous iteration. Constraint (27)
a part of the proposed BDMDS model. To address this issue, a only limits the sum of output power in three time intervals that
novel algorithm, neural network and optimization decoupled have the highest battery charging or discharging power, named
heuristic, is proposed in this paper to decouple the complex as precise battery consumption limit (PBCL). Constraints
BDMDS model and solve it in the following five steps in an (28)-(29) are designed to limit the maximum charg-
iterative manner. ing/discharging power that is named as battery C rate limit
• Step A is to solve the microgrid day-ahead scheduling (BRL). The proposed constraints will be generated and updat-
with additional constraints by limiting BESS usage that ed in each iteration by using the scheduled battery operational
is generated from Step E. Note that in the first iteration, profile from previous iteration and the proposed battery opera-
there is no extra limit on BESS usage. tion restriction factor (BORF) 𝛼𝛼 which is a preset parameter.
• Step B obtains the scheduled BESS operating profile Note that the value of BORF can determine the limits in BCL,
that is then processed with the proposed CBUP method. PBCL, and BRL for each iteration.
𝑡𝑡,𝑖𝑖 𝑡𝑡,𝑖𝑖 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃_𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆
• Step C estimates BESS degradation with the proposed �(𝑃𝑃𝐶𝐶ℎ𝑎𝑎𝑎𝑎 + 𝑃𝑃𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 ) ≤ (1 − 𝛼𝛼) ∗ 𝑃𝑃𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 (26)
NNBD model that is trained beforehand and calculates 𝑡𝑡∈𝑇𝑇
the associated equivalent battery degradation cost. 𝑡𝑡,𝑖𝑖
� 𝑃𝑃𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷
𝑡𝑡, 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃_𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆
+ 𝑃𝑃𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 ≤ (1 − 𝛼𝛼) ∗ 𝑃𝑃𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 (27)
• Step D determines whether the solutions meet the de- 𝑡𝑡∈𝑇𝑇𝑇𝑇𝑇𝑇 3
signed stopping criteria: stop the iteration process and 𝑡𝑡,𝑖𝑖
𝑃𝑃𝐶𝐶ℎ𝑎𝑎𝑎𝑎 ≤ 𝑃𝑃𝑆𝑆𝑆𝑆𝑀𝑀𝑀𝑀𝑀𝑀 (1 − 𝛼𝛼)𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖−1 (28)
report the solution if yes; otherwise, go to Step E. 𝑡𝑡,𝑖𝑖 𝑀𝑀𝑀𝑀𝑀𝑀
𝑃𝑃𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 ≤ 𝑃𝑃𝑆𝑆𝑆𝑆 (1 − 𝛼𝛼) 𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖−1 (29)
• Step E updates the boundaries for the BESS related op-
erating constraints from section IV.B which can limit Depending on which battery usage limiting constraint is
the BESS operations to reduce the degradation cost; used in CMDS, there are four variations of the proposed
and the associated constraints will be sent to Step A to NNODH algorithm. They are all conducted to compare the
be included in MDS for the next iteration. performance of the proposed constraints. The proposed four
variations are presented in Table III. The CMDS model in this
This iterative procedure is illustrated in Fig. 3. table is solved instead of the traditional MDS model starting
from the second iteration of the proposed NNODH algorithm.
IEEE Transactions on Power Systems 7

Table III The proposed four NNODH strategies cost reduction (TCR) and operational cost increment (OCI).
Strategies CMDS Model for Step A They are defined in (30), (31), and (32) respectively. The deg-
NNODH-BCL (1)-(15), (26)
NNODH-PBCL (1)-(15), (27)
radation cost reduction illustrates how much the battery deg-
NNODH-BRL (1)-(15), (28)-(29) radation cost can be reduced as compared to the maximum
NNODH-ALL (1)-(15), (26)-(29) value when battery degradation is not considered in traditional
C. Iterations and Stopping Criteria MDS. The TCR represents how much the total cost can be
reduced by the proposed model while the OCI shows how
Fig. 4 illustrates the iteration process of determining the op- much the operation cost would increase when considering the
timal solution for the battery degradation based MDS using
battery degradation model. Notations 𝐵𝐵𝐵𝐵𝐵𝐵 𝑀𝑀𝑀𝑀𝑀𝑀 , 𝑇𝑇𝑇𝑇 𝑀𝑀𝑀𝑀𝑀𝑀 , 𝑂𝑂𝑂𝑂 𝑀𝑀𝑀𝑀𝑀𝑀
the proposed NNODH algorithm. As is discussed in Section
represent the maximum battery degradation cost, maximum
IV.A, the first iteration only considers the traditional mi-
total cost, and the minimum operation cost respectively. These
crogrid day-ahead scheduling model that ignores the battery
metrics are obtained by solving the traditional MDS problem
degradation. The NNODH algorithm can assure that the solu-
in the first iteration when battery degradation cost is not con-
tion is guaranteed to reach optimal. The first iteration will
sidered.
provide the optimal solution when the battery degradation is 𝐵𝐵𝐵𝐵𝐵𝐵 𝑀𝑀𝑀𝑀𝑀𝑀 −𝐵𝐵𝐵𝐵𝐵𝐵
not considered. The BESS operation will be compressed or 𝐷𝐷𝐷𝐷𝐷𝐷 = ∗ 100% (30)
𝐵𝐵𝐵𝐵𝐵𝐵 𝑀𝑀𝑀𝑀𝑀𝑀
limited to reduce the battery degradation in the rest of the iter- 𝑇𝑇𝑇𝑇 𝑀𝑀𝑀𝑀𝑀𝑀 −𝑇𝑇𝑇𝑇
𝑇𝑇𝑇𝑇𝑇𝑇 = ∗ 100% (31)
ations. During the iterations, the battery degradation will de- 𝑇𝑇𝑇𝑇 𝑀𝑀𝑀𝑀𝑀𝑀
crease gradually while the operation cost will increase gradu- 𝑂𝑂𝑂𝑂−𝑂𝑂𝑂𝑂 𝑀𝑀𝑀𝑀𝑀𝑀
𝑂𝑂𝑂𝑂𝑂𝑂 = ∗ 100% (32)
𝑂𝑂𝑂𝑂 𝑀𝑀𝑀𝑀𝑛𝑛
ally. Thus, the vertex point of the total cost curve is the opti-
When we consider these proposed metrics to evaluate the
mal solution. In Fig. 4, the red star denotes the global optimal
proposed algorithm, the TCR will be the first choice since the
solution for the battery degradation based MDS that considers
objective of the algorithm is to reduce the total cost. The algo-
the proposed NNBD model; and the black star represents the
rithm with a higher TCR is always preferred in the analysis. In
solution for MDS/CMDS problem in the current iteration
the meanwhile, the DCR and OCI are designed to gauge the
while the grey star denotes the solution for the MDS/CMDS
degradation cost and operation cost. We prefer a higher num-
problem in the previous iteration. The MDS/CMDS solutions
ber of the DCR and a lower number of the OCI. However,
are approximations to the global optimal solution for the bat-
DCR and OCI are not as important as the TCR here.
tery degradation based MDS; they are referred to as pseudo
A stopping criterion is proposed as part of the NNODH al-
solutions (feasible but may not be global optimal) for the bat-
gorithm. It is used to terminate the iterations and then the best
tery degradation based MDS. Graph 1 in Fig. 4 represents the
solution can be reported. The designed stopping criteria is
first iteration. The BESS usage limit constraints that generated
defined as follows: for any past 10 iterations, if the 𝑇𝑇𝑇𝑇 𝑛𝑛 –
in each iteration will reduce the feasible region for the
𝑇𝑇𝑇𝑇 𝑛𝑛−1 is less than 0 for the first five iterations but greater than
MDS/CMDS problem in the next iteration. Graphs 2-4 show
0 for the last five iterations, then the iteration will stop and the
the feasible region shrinks after each iteration and the black
best solution with lowest total cost for BDMDS will be report-
star (pseudo solution) is moving towards the red star (global
ed. Here, 𝑇𝑇𝑇𝑇 𝑛𝑛 and 𝑇𝑇𝑇𝑇 𝑛𝑛−1 denotes the total costs in nth and (n-
optimal solution for BDMDS). Graph 5 shows the red star and
1)th iterations respectively. The total cost here represents the
black star overlap which indicates that the black star here is
lowest total cost rather than the lowest battery degradation
the optimal solution for the BDMDS problem. If the iteration
cost. The optimal solution may be affected by the battery size
continues, the feasible solution area of BDMDS will continue
as well as the unit price.
to reduce such that the optimal solution might be cut out.
Since the proposed NNODH algorithm is an iterative method, E. Benchmark Models of MDS
the 𝛼𝛼 values (BORF) in (26)-(29) would change over itera- Benchmark models are designed to evaluate and demon-
tions and may result into different BDMDS solutions, among strate the effectiveness and performance of the proposed
which the best solution should deviate slightly away from the NNODH algorithm. Three benchmark models are presented
global optimal solution. To ensure the best pseudo solution is below and summarized in Table IV:
close enough to the global solution with least total cost, small 1. Traditional MDS Model: The basic benchmark model is
𝛼𝛼 value is preferred. However, it may take more iterations for set with no BESS degradation cost and the MDS will
a smaller 𝛼𝛼 value to find the optimal solution. maximize the usage of BESS to minimize the operation
cost of the microgrid.
2. Cycle Limit Model: This model is widely adopted in the
industry. It is designed to limit the charging and discharg-
ing cycles to decrease the degradation of the BESS. This
model is formulated with the following constraints where
𝑡𝑡,𝑖𝑖
the 𝑉𝑉𝐶𝐶ℎ𝑎𝑎𝑎𝑎/𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 is a binary variable that represents the sta-
tus change of charging or discharging operation.
Fig. 4. Illustration of the proposed NNODH algorithm.
𝑡𝑡,𝑖𝑖
𝑉𝑉𝐶𝐶ℎ𝑎𝑎𝑎𝑎/𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 𝑡𝑡,𝑖𝑖
≤ 𝑈𝑈𝐶𝐶ℎ𝑎𝑎𝑎𝑎/𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 𝑡𝑡−1,𝑖𝑖
+ 𝑈𝑈𝐶𝐶ℎ𝑎𝑎𝑎𝑎/𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 , ∀𝑖𝑖, 𝑡𝑡 (33)
D. Evaluation Metrics
Three metrics are designed to evaluate the performance of
𝑡𝑡,𝑖𝑖
𝑉𝑉𝐶𝐶ℎ𝑎𝑎𝑎𝑎/𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 ≥ 𝑡𝑡,𝑖𝑖
𝑈𝑈𝐶𝐶ℎ𝑎𝑎𝑎𝑎/𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 − 𝑡𝑡−1,𝑖𝑖
𝑈𝑈𝐶𝐶ℎ𝑎𝑎𝑎𝑎/𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 , ∀𝑖𝑖, 𝑡𝑡 (34)
the proposed model: degradation cost reduction (DCR), total 𝑡𝑡,𝑖𝑖
𝑉𝑉𝐶𝐶ℎ𝑎𝑎𝑎𝑎/𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 ≥ 𝑡𝑡−1,𝑖𝑖
𝑈𝑈𝐶𝐶ℎ𝑎𝑎𝑎𝑎/𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 − 𝑡𝑡,𝑖𝑖
𝑈𝑈𝐶𝐶ℎ𝑎𝑎𝑎𝑎/𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 , ∀𝑖𝑖, 𝑡𝑡 (35)
IEEE Transactions on Power Systems 8

𝑡𝑡,𝑖𝑖
𝑉𝑉𝐶𝐶ℎ𝑎𝑎𝑎𝑎/𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 𝑡𝑡,𝑖𝑖
≤ 2 − 𝑈𝑈𝐶𝐶ℎ𝑎𝑎𝑎𝑎/𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 𝑡𝑡−1,𝑖𝑖
− 𝑈𝑈𝐶𝐶ℎ𝑎𝑎𝑎𝑎/𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 , ∀𝑖𝑖, 𝑡𝑡 (36) hours, which is solved by the Pyomo package [41] with the
1
𝑉𝑉𝑐𝑐/𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑 = 0 (37) Gurobi solver [42].
𝑡𝑡,𝑖𝑖 B. Training Results of NNBD
� 𝑉𝑉𝐶𝐶ℎ𝑎𝑎𝑎𝑎/𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 ≤2 (38)
𝑡𝑡
The training results of NNBD and hyperparameters are pre-
3. Linear Battery Degradation Cost (BDC) Model: The third sented and analyzed in this section. Mini-batch technology is
benchmark model is based on constant battery degrada- used for the NN training. The optimal batch size may vary for
tion cost parameters. It is assumed that the battery degra- different input data. Different training batch sizes are tested to
dation cost is linear to the energy consumption of the determine the optimal training batch-size and the test results
BESS as shown in (39) where the 𝑐𝑐𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 is a fixed rate are presented in Table V. It can be observed that the batch size
that represents the battery degradation cost per unit usage of 256 can achieve the highest accuracy while requiring less
of BESS. epochs to complete the training process. The validation accu-
racy can reach up to 94.5% while it only requires 50 epochs to
𝑡𝑡,𝑖𝑖 reach a steady accuracy. The error tolerance is set to 15%
𝑓𝑓 𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 = 𝑐𝑐𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 � 𝑃𝑃𝐶𝐶ℎ𝑎𝑎𝑎𝑎/𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷 ∗ 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙ℎ(𝑡𝑡) (39) when calculating prediction accuracy in Table V and Table
𝑡𝑡,𝑖𝑖
VI. Larger or smaller batch size will either reduce the accura-
Table IV Benchmark models
MDS Models Formulations cy or increase the training epochs. We found that large batch
Traditional MDS (1)-(15) size can help smooth the oscillation of the training accuracy
Cycle Limit (1)-(15), (33-38) curve. The training accuracies with batch sizes of 128 and 256
Linear BDC (1)-(15), (39) are almost equally the best. However, the training accuracy
curve with a batch size of 256 is much smoother than the other
V. CASE STUDIES one. Moreover, we observed that if the input data are shuffled,
A. Microgrid Testbed then the neural network cannot obtain good results. This may
A typical grid-connected microgrid with renewable energy be due to the characteristic of the input data: the battery deg-
sources is created in this paper as a testbed to examine the radation data are time-series for each battery aging test. The
performance of the proposed NNBD model, BDMDS model results with different data pre-processing methods are com-
and NNODH algorithm. This testbed includes one 180kW pared in Table VI. Note that the test results in Table VI are
diesel generator (DG), five 200kW wind turbines (WT), 300 obtained with the batch-size of 256 for all the trainings. The
residential houses that contains solar panel (5kW capacity per regressed data pre-processing method has the highest accuracy
house), and a 300kWh lithium-ion based BESS with a charg- and efficiency. Based on the results from Table V and Table
ing/discharging efficiency of 90%. The load data representing VI, the regressed method performs the best and is applied for
1000 residential houses. The ambient temperature and availa- all subsequent simulations.
ble solar power for a time period of 24 hours are obtained
from the Pecan Street Dataport [39]. The wholesale electricity
price is obtained from ECROT [40]. The price of the electrici-
ty sold to the main grid is set to 80% of the purchase price.
Sensitivity analysis is conducted with different RES penetra-
tion levels and different BESS sizes. The computer with In-
tel® Xeon(R) W-2295 CPU @ 3.0 GHz, 256 GB of RAM,
and Nvidia Quadro RTX 8000 (48GB GPU) was utilized to
conduct the numerical simulations including the training of
battery degradation model and the optimization of microgrid
day-ahead scheduling. The microgrid resource scheduling
problem studied in this work covers a total time horizon of 24
Fig. 5. DNN training and validation results.
Table V Sensitivity test of batch-sizes
Batch-Size 16 32 64 128 256 512 1024 2048
Number of Epochs 20 20 85 110 50 50 75 75
Accuracy 40% 50% 90.5% 94.0% 94.5% 92.5% 88% 67%
Table VI Training results with different data pre-processing methods 20% error tolerance is 95.5% which is only 1% higher than the
Data Pre-processing Number of 15% error tolerance. Thus, we choose 15% as the error toler-
Accuracy
Methods Epochs ance level when calculating accuracy in all subsequent results.
Raw 78% 150 After 50 epochs, the training loss and the validation loss stop
Smoothed 82% 100
Regressed 94.5% 50
decreasing and the training accuracy becomes stable as well.
The training stops at 65th epoch to avoid overfitting and oscil-
The best training results are shown in Fig. 5 that illustrates lation.
the training loss versus validation loss and the accuracy curves
under different error thresholds. The accuracy is around 60% C. Results of NNODH algorithm
under a 5% error tolerance, 80% under a 10% error tolerance In this section, the BDMDS results obtained with the
and 94.5% under a 15% error tolerance. The accuracy with NNODH algorithm are presented. The load profile of the test
IEEE Transactions on Power Systems 9

bed is shown in Fig. 6. Table VII presents the results for dif- The scheduled BESS operation for different models are
ferent strategies of the proposed NNODH algorithm presented shown in Fig. 8. Positive output means the BESS is in dis-
in Section IV. The BCL, PBCL, BRL and ALL CMDS models charging mode and negative output means it is in charging
are implemented and tested separately. The combination of all mode. The scheduled BESS operations for the Traditional
three types of constraints as the added constraints to MDS in MDS model, Cycle Limit model, Linear BDC model and the
the next iteration, marked as ALL in Table VII, is tested as proposed BDMDS model are all shown in Fig. 8. It can be
well. In Table VII, for all the proposed models, the initial iter- observed for the traditional MDS that does not consider the
ation does not have any limits of the BESS operation. This battery degradation, BESS operates at a wider output range
also leads to the solution of the first iteration having the high- from -150 kW to 150 kW in seven different time intervals.
est battery degradation cost. Similarly, the maximum total cost When battery degradation is considered, BESS is scheduled to
for different models is the same. For metric DCR, the ALL charge and discharge in a narrow range and in less active time
option performs the best among all strategies, and it decreases intervals. The BESS operation patterns for different models
the battery degradation cost by 79.27%. BCL performs the match in most of the time intervals, which proves the effec-
best in terms of metric TCR. The increased operation cost is tiveness of those four models. For the proposed BDMDS, after
similar between BCL and PBCL, which is less than BRL and applying the NNODH algorithm to solve it, the BESS operates
ALL. Overall, we prefer the NNODH-BCL strategy due to its only in three different time intervals. Total exchanged energy
best performance to decrease the total cost and the fast solving is limited to reduce the battery degradation which meets the
time. designed purpose of constraint BCL. The total usage of BESS
for Traditional MDS model is 920 kWh while it is 325 kWh
2000 for BDMDS, 601 kWh for Cycle Limit model and 473 kWh
1500 for Linear BDC model.
Load/Generation [KW]

1000

500

0
1 3 5 7 9 11 13 15 17 19 21 23
-500

-1000
Load Net-Load Renewable
-1500
Time [h]
Fig. 6. Load profile of the microgrid testbed.
Table VII Results for proposed strategies
NNODH BCL PBCL BRL ALL
Total Cost ($) 494.36 503.49 502.50 500.32
Degradation Cost ($) 10.74 19.50 11.99 10.39 Fig. 7. BDMDS Results of the NNODH-BCL method.
TCR 5.82% 4.08% 4.27% 4.69%
Traditional MDS
DCR 78.57% 60.23% 76.08% 79.27%
OCI 1.83% 1.81% 3.20% 3.09% 140 BDMDS-BCL
Solving time (s) 8.68 12.82 10.00 8.48 Cycle Limit
BESS Output Power \kW

90
Iteration Numbers 37 56 46 37
Linear BDC
Fig. 7 shows the results for NNODH-BCL including the 40
degradation cost, operation cost, battery degradation in per-
-10
cent, and total cost. These results are based on the battery unit 1 2 3 4 5 6 7 8 9 101112131415161718192021222324
price of 400 $/kWh and the value of BORF is set to 0.03. -60
From Fig. 7, we can observe that the valley of the total cost
curve is at the 37th iteration with $494.36 including the equiva- -110
lent battery degradation cost of $10.74 and the microgrid op-
-160
eration cost of $483.62. The total cost is reduced by 5.82% Hour
compared to the traditional MDS model that does not consider Fig. 8. BESS scheduled operations comparison.
the battery degradation cost, which proves that the proposed
NNODH algorithm can reduce the total cost significantly. Table VIII presents the results of the CBUP method applied
Note that the results shown in Fig. 7 did not implement the on the solutions obtained from BDMDS model and traditional
stopping criteria to show how the system behaves when the MDS model. It can be clearly observed that the battery degra-
battery’s usage is further limited until idle. If not, the iteration dation prediction is pretty close to the true battery degradation
will stop at the 42nd iteration since the minimum total cost can in a small percent of error. Without the CBUP method, the
be found at the 37th iteration. After the 37th iteration, we can error of BD prediction from NNBD model is significantly
observe that even though the battery degradation cost keeps higher. Thus, the results proves the effectiveness of the pro-
decreasing, the total cost starts to increase due to the increased posed CBUP method. From Table IX, we can observe that the
slope of the operation cost. proposed BDMDS model has the lowest daily degradation and
IEEE Transactions on Power Systems 10

the lowest annual degradation cost among all the models. higher total cost reduction in percent. For the same unit price
Moreover, the BDMDS model is outstanding in annual cost of BESS, higher BESS size tends to achieve a lower TCR.
saving. Also, if we consider the load profile is the same for Table XII shows that for the same unit price except
each day, then the proposed BDMDS model can significantly $200/kWh, higher BESS size has a lower DCR. This may be
extend the lifetime of the BESS than all the benchmark mod- because that for the same battery output power, bigger battery
els. Note that the expect end of life is set to 70% of SOH val- size will likely have a lower C rate which is one of the main
ue. contributing factors of battery degradation. Also, for the same
Table X shows the results of the sensitivity tests with dif- size BESS, higher unit price lead to a higher DCR. Thus, we
ferent BORF values. The results show that larger BORF can can conclude that the BESS price of per unit capacity is the
achieve the lowest total cost in less iterations and less solving major factor affecting how much the proposed model reduces
time. As a tradeoff, the optimal total cost of larger BORF will the degradation cost. The BESS size is the major factor affect-
be slightly higher. Smaller BORF will lead to a smaller area ing how much the proposed model reduces the battery degra-
cut from the feasible solution area in each iteration, which dation. The result of the sensitivity analysis in this section may
requires more iterations to converge to the optimal solution. help determine the size of the BESS for microgrid planning. In
However, the optimal solutions for the different BORFs do not the meanwhile, as the unit price of BESS keeps decreasing,
have a significant difference. Thus, a higher value of BORF the battery degradation cost will be lower and account for
such as 0.05 is preferred due to the high computing efficiency. smaller percentage respect to the total cost in the future.
Table VIII Battery degradation prediction with different methods 7%
Without CBUP With CBUP 200 kWh 300 kWh 400 kWh
Model True BD 6%
Degradation Error Degradation Error
BDMDS 0.00468% 0.037% 690% 0.0044% 5.9% 5% 6.09%
MDS 0.0108% 0.066% 511% 0.0120% 9.7%
4% 5.09% 5.82%

TCR
Table IX Model comparison 4.04%
3% 4.94%
Annual Annual Expect 3.13%
Daily BESS 2% 4.09%
Model Degradation Cost Lifetime 3.69%
Degradation 2.20% 3.09% 3.03%
Cost ($) Saving ($) (years) 2.43%
1% 2.11% 1.85%
MDS 0.02% 18,301.1 N/A 4.1 1.26%
Cycle Limit 0.012% 12,540.8 6,205 6.8 0%
Linear BDC 0.01% 8,832.5 6,935 8.2 200$/kWh 250$/kWh 300$/kWh 350$/kWh 400$/kWh
BDMDS 0.0045% 3,920.1 11,151 18.3 Unit Cost of BESS
Fig. 9. Results of TCR sensitivity tests.
Table X Results of sensitivity analysis with different BORF values
BORF Number of Iterations Optimal Total Cost ($) Time (s)
Table XII DCR of different BESS sizes and unit prices
0.01 113 494.35 25.21
0.02 57 494.35 12.51 Unit Price ($/kWh)
0.03 38 494.36 8.42 Size (kWh) 200 250 300 350 400
0.05 22 494.36 5.20 200 75.15% 82.16% 82.96% 85.18% 88.58%
0.1 11 494.38 2.85 300 77.29% 77.29% 78.57% 78.57% 78.57%
0.2 5 494.40 1.41 400 52.95% 54.48% 54% 71.90% 72.68%

D. Sensitivity Analysis of RES Penetration Levels VI. CONCLUSIONS

Table XI presents the TCR and OCI results under different In this paper, a neural network based battery degradation
RES penetration levels for the microgrid. The penetration lev- model is proposed to predict the BESS degradation value for
el is defined as the ratio of the average renewable generation each scheduling period. A cycle based battery usage pro-
to the average load in a typical day. The proposed microgrid cessing method is designed to accurately apply the proposed
testbed is set to an 80% RES penetration level. From Table XI, NNBD model with the microgrid day-ahead scheduling mod-
we can observe that with the increase of RES penetration lev- el. The NNODH algorithm is proposed to decouple the battery
el, the value of TCR will increase and the value of OCI will degradation based microgrid day-ahead scheduling problem
decrease. This is because that in a system with higher RES that is hard to solve directly due to the highly non-linear char-
penetration, BESS is required to charge/discharge more fre- acteristic of the proposed NNBD model. The proposed
quently to mitigate the uncertainty resulted by the RES gen- NNODH algorithm can solve the MDS optimization problem
eration, which leads to a higher battery degradation. and calculate the battery degradation cost iteratively and effec-
Table XI Results of different RES penetration levels tively find the optimal solution with the lowest total cost. The
RES Penetration Level CMDS problem is created with the options of using different
20% 40% 60% 80% proposed constraints to limit the BESS usage and obtain a
TCR 1.60% 2.1% 3.12% 5.82% lower battery degradation cost. An RES-enriched microgrid is
OCI 0.47% 0.63% 0.94% 1.83% used to evaluate the performance of the proposed BDMDS
E. Sensitivity Analysis of BESS Unit Prices and Sizes model and the proposed NNODH algorithm.
The test results demonstrate that the battery degradation can
In this section, the sensitivity analysis on different BESS
be accurately predicted (5.9% error) by the proposed NNBD
unit prices and sizes is conducted. Fig. 9 shows that the total
model with the adjusted inputs obtained by the proposed
cost reduction in percentage. It can be clearly observed that for
CBUP method. The proposed NNODH algorithm can obtain
the same size of BESS, higher unit price corresponds to a
the optimal solution efficiently. Compared with the traditional
IEEE Transactions on Power Systems 11

MDS models, the total cost can be reduced significantly by [11] Z. Miao, L. Xu, V. R. Disfani and L. Fan, "An SOC-Based Battery
Management System for Microgrids," in IEEE Transactions on Smart
5.82% with the proposed BDMDS model. Also, the proposed Grid, vol. 5, no. 2, pp. 966-973, March 2014.
model can reduce the daily BESS degradation significantly [12] B. Li, T. Chen, X. Wang and G. B. Giannakis, "Real-Time Energy
from 0.02% to 0.0045%. The annual degradation cost is re- Management in Microgrids With Reduced Battery Capacity Require-
duced by 78.6% with the proposed model. Moreover, the ex- ments," in IEEE Transactions on Smart Grid, vol. 10, no. 2, pp. 1928-
1938, March 2019.
pected lifetime is extended from 4.1 years to 18.3 years with [13] C. Zhao and X. Li,“A Novel Real-Time Energy Management Strategy
the proposed model. The NNODH-BCL performs the best for Grid-Friendly Microgrid: Harnessing Internal Fluctuation Internal-
among the proposed four strategies in this case. Sensitivity ly”, 52nd North American Power Symposium, (Virtually), Tempe, AZ,
tests demonstrate the performance of the proposed NNODH USA Apr. 2021.
[14] C. Zhao and X. Li,“A Novel Real-Time Energy Management Strategy
algorithm under different BESS sizes and unit prices. Overall, for Grid-Supporting Microgrid: Enabling Flexible Trading Power”,
this work demonstrates the effectiveness of the proposed IEEE PES General Meeting, (Virtually), Washington D.C., USA, Jul.
BDMDS model for reducing battery degradation cost and total 2021.
cost, and the capability of the proposed NNODH algorithm for [15] C. Ju, P. Wang, L. Goel and Y. Xu, "A Two-Layer Energy Manage-
ment System for Microgrids With Hybrid Energy Storage Considering
efficiently solving BDMDS that is a deep neural network em- Degradation Costs," in IEEE Transactions on Smart Grid, vol. 9, no.
bedded optimization problem. 6, pp. 6047-6057, Nov. 2018.
[16] B. Foggo and N. Yu, "Improved Battery Storage Valuation Through
VII. FUTURE WORK Degradation Reduction," in IEEE Transactions on Smart Grid, vol. 9,
no. 6, pp. 5721-5732, Nov. 2018.
The uncertainty of the renewable generation can be consid- [17] L. Zhang et al., "Improved Cycle Aging Cost Model for Battery Ener-
ered and addressed with advanced models such as a stochastic gy Storage Systems Considering More Accurate Battery Life Degra-
and robust optimization model, which is a potential extension dation," in IEEE Access, vol. 10, pp. 297-307, 2022.
of this work. Moreover, battery aging test data are often very [18] M. Sandelic, A. Sangwongwanich and F. Blaabjerg, "Incremental
Degradation Estimation Method for Online Assessment of Battery
limited, not supporting deep learning studies that require large Operation Cost," in IEEE Transactions on Power Electronics, vol. 37,
amounts of data. One solution we plan to develop to address no. 10, pp. 11497-11501, Oct. 2022.
this issue is to leverage transfer learning [43] with the trained [19] Y Shi et al., "A Convex Cycle-based Degradation Model for Battery
NNBD model in this paper to obtain different NN models for Energy Storage Planning and Operation", 2018 Annual American
Control Conference (ACC), 2018.L.
other types of batteries. This would enable us to train battery [20] De Pascali, F. Biral and S. Onori, "Aging-Aware Optimal Energy
degradation models for different types of batteries with much Management Control for a Parallel Hybrid Vehicle Based on Electro-
less data. chemical-Degradation Dynamics," in IEEE Transactions on Vehicular
Technology, vol. 69, no. 10, pp. 10868-10878, Oct. 2020.
REFERENCES [21] J. Cao, D. Harrold, Z. Fan, T. Morstyn, D. Healey and K. Li, "Deep
Reinforcement Learning-Based Energy Storage Arbitrage With Accu-
[1] B. Li, T.Chen, X. Wang and G. Giannakis, “Real-Time Energy Man- rate Lithium-Ion Battery Degradation Model," in IEEE Transactions
agement in Microgrids With Reduced Battery Capacity Require- on Smart Grid, vol. 11, no. 5, pp. 4513-4521, Sept. 2020.
ments,” IEEE Trans. Smart Grid., vol. 10, no. 2, pp. 1928–1938, Mar. [22] G. Abdelaal, M. I. Gilany, M. Elshahed, H. M. Sharaf and A.
2019. El’gharably, "Integration of Electric Vehicles in Home Energy Man-
[2] M. T. Lawder et al., "Battery Energy Storage System (BESS) and Bat- agement Considering Urgent Charging and Battery Degradation," in
tery Management System (BMS) for Grid-Scale Applications," in IEEE Access, vol. 9, pp. 47713-47730, 2021.
Proceedings of the IEEE, vol. 102, no. 6, pp. 1014-1030, June 2014. [23] K. Abdulla et al., "Optimal operation of energy storage systems con-
[3] X. Gong, R. Xiong and C. C. Mi, "Study of the Characteristics of Bat- sidering forecasts and battery degradation," 2017 IEEE Power & En-
tery Packs in Electric Vehicles With Parallel-Connected Lithium-Ion ergy Society General Meeting, 2017, pp. 1-1.
Battery Cells," in IEEE Transactions on Industry Applications, vol. [24] K. Antoniadou-Plytaria, D. Steen, L. A. Tuan, O. Carlson and M. A.
51, no. 2, pp. 1872-1879, March-April 2015. Fotouhi Ghazvini, "Market-Based Energy Management Model of a
[4] Z. Lyu, R. Gao, and L. Chen, “Li-Ion Battery State of Health Estima- Building Microgrid Considering Battery Degradation," in IEEE
tion and Remaining Useful Life Prediction Through a Model-Data- Transactions on Smart Grid, vol. 12, no. 2, pp. 1794-1804, March
Fusion Method,” IEEE TRANSACTIONS ON POWER 2021.
ELECTRONICS., vol. 36, no. 6, pp. 6228–6240, June. 2021. [25] G. Saldaña, J. I. S. Martín, I. Zamora, F. J. Asensio, O. Oñederra and
[5] B. Xu, A. Oudalov, A. Ulbig, G. Andersson and D. Kirschen, "Model- M. González, "Empirical Electrical and Degradation Model for Elec-
ing of lithium-ion battery degradationfor cell life assessment," 2017 tric Vehicle Batteries," in IEEE Access, vol. 8, pp. 155576-155589,
IEEE Power & Energy Society General Meeting, 2017, pp. 1-1. 2020.
[6] X. Han, L. Lu, Y. Zheng, X. Feng, Z. Li, J. Li, M. Ouyang, “A review [26] Z. Wei, Y. Li and L. Cai, "Electric Vehicle Charging Scheme for a
of the key issues of the lithium ion battery degradation among the Park-and-Charge System Considering Battery Degradation Costs," in
whole life cycle”, in eTransportation, vol 1, Jul. 2019. IEEE Transactions on Intelligent Vehicles, vol. 3, no. 3, pp. 361-373,
[7] P. S. Kumar, R. P. S. Chandrasena, V. Ramu, G. N. Srinivas and K. V. Sept. 2018.
S. M. Babu, "Energy Management System for Small Scale Hybrid [27] G. Saldaña, J. I. S. Martín, I. Zamora, F. J. Asensio, O. Oñederra and
Wind Solar Battery Based Microgrid," in IEEE Access, vol. 8, pp. M. González, "Empirical Electrical and Degradation Model for Elec-
8336-8345, 2020. tric Vehicle Batteries," in IEEE Access, vol. 8, pp. 155576-155589,
[8] A. Merabet, K. Tawfique Ahmed, H. Ibrahim, R. Beguenane and A. 2020.
M. Y. M. Ghias, "Energy Management and Control System for Labor- [28] Y. Sun, H. Yue, J. Zhang and C. Booth, "Minimization of Residential
atory Scale Microgrid Based Wind-PV-Battery," in IEEE Transac- Energy Cost Considering Energy Storage System and EV With Driv-
tions on Sustainable Energy, vol. 8, no. 1, pp. 145-154, Jan. 2017. ing Usage Probabilities," in IEEE Transactions on Sustainable Ener-
[9] X. Xing, L. Xie, H. Meng, X. Guo, L. Yue and J. M. Guerrero, "Ener- gy, vol. 10, no. 4, pp. 1752-1763, Oct. 2019.
gy management strategy considering multi-time-scale operational [29] H. Farzin, M. Fotuhi-Firuzabad and M. Moeini-Aghtaie, "A Practical
modes of batteries for the grid-connected microgrids community," in Scheme to Involve Degradation Cost of Lithium-Ion Batteries in Ve-
CSEE Journal of Power and Energy Systems, vol. 6, no. 1, pp. 111- hicle-to-Grid Applications," in IEEE Transactions on Sustainable En-
121, March 2020. ergy, vol. 7, no. 4, pp. 1730-1738, Oct. 2016.
[10] K. Thirugnanam, S. K. Kerk, C. Yuen, N. Liu and M. Zhang, "Energy [30] Z. Wei, Y. Li and L. Cai, "Electric Vehicle Charging Scheme for a
Management for Renewable Microgrid in Reducing Diesel Generators Park-and-Charge System Considering Battery Degradation Costs," in
Usage With Multiple Types of Battery," in IEEE Transactions on In-
dustrial Electronics, vol. 65, no. 8, pp. 6772-6786, Aug. 2018.
IEEE Transactions on Power Systems 12

IEEE Transactions on Intelligent Vehicles, vol. 3, no. 3, pp. 361-373, [36] S. Saxena, C. Floth, J. MacDonald. S. Moura, "Quantifying EV bat-
Sept. 2018. tery end-of-life through analysis of travel needs with vehicle power-
[31] S. Fang, B. Gou, Y. Wang, Y. Xu, C. Shang and H. Wang, "Optimal train models," in Journal of Power Sources, vol. 282, pp. 265-276.
Hierarchical Management of Shipboard Multibattery Energy Storage Jan. 2015.
System Using a Data-Driven Degradation Model," in IEEE Transac- [37] S. Loffe, C.Szegedy, "Batch Normalization: Accelerating Deep Net-
tions on Transportation Electrification, vol. 5, no. 4, pp. 1306-1318, work Training by Reducing Internal Covariate Shift," in Proceedings
Dec. 2019. of the 32nd International Conference on Machine Learning, Vol.
[32] P. Attia, M. Deetjen, and J. Witmer. “Accelerating battery develop- 37,pp. 448-456, Mar. 2015.
ment by early prediction of cell lifetime,” (2018). [38] NN-SVG, http://alexlenail.me/NN-SVG/index.html.
[33] P. Attia, A. Grover, N. Jin et al., “Closed-loop optimization of fast- [39] "Dataport Resources," May, 2019 [online] Available:
charging protocols for batteries with machine learning,” in Na- https://dataport.pecanstreet.org/academic.
ture, vol. 578, pp. 397–402, 2020. [40] "ERCOT, Electric Reliability Council of Texas," [Online]. Available:
[34] K. Severson, P. Attia, N. Jin, et al., “Data-driven prediction of battery http://www.ercot.com/.
cycle life before capacity degradation,” in Nat Energy vol. 4, pp 383– [41] "Pyomo, Python Software packages.," Available: [Online]. Available:
391, 2019. http://www.pyomo.org/ .
[35] Omar N., M. A. Monem, Y. Firouz, J. Salminen, J. Smekens, O. He- [42] "Gurobi Optimization, Linear Programming Solver," [Online]. Avail-
gazy, H. Gaulous, G. Mulder, P. Van den Bossche, T. Coosemans, and able: https://www.gurobi.com/.
J. Van Mierlo. “Lithium iron phosphate based battery — Assessment [43] L. Shao, F. Zhu and X. Li, "Transfer Learning for Visual Categoriza-
of the aging parameters and development of cycle life model.” Ap- tion: A Survey," in IEEE Transactions on Neural Networks and
plied Energy, Vol. 113, January 2014, pp. 1575–1585. Learning Systems, vol. 26, no. 5, pp. 1019-1034, May 2015.