Академический Документы
Профессиональный Документы
Культура Документы
net/publication/328146917
CITATIONS READS
0 23
4 authors, including:
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Mehdi Assefi on 16 October 2018.
Inverter Controller
PV
DC
AC
Meter
Fig. 2: A behind-the-meter (BTM) architecture utilizing PV Fig. 3: An example of daily battery operation for the applica-
and storage solutions. tion of reducing demand charge (DC).
pacity fading and significantly outperforms baseline models, discharging scheduling and hence minimize the demand charge
including persistence and autoregressive (AR) models. in the electricity bill [15]. In other words, lack of modeling
The rest of this paper is organized as follows. Section II the capacity degradation or an inaccurate fading model causes
provides a background on use cases of capacity devices in the scheduling module to take into account a wrong capacity
power system, and Section III reviews the related works. value which results in suboptimal scheduling of battery usage.
Section IV describes the battery operation data used to develop This in turn increases the peak demand charge and incurs costs
the model and provides the preprocessing techniques applied to users.
on the data. In section V, we propose the deep learning-based In the EMS architecture designed in Fig. 1, capacity es-
method for modeling the capacity fading. Section VI provides timation will be used to adjust the available capacity for
the experimental results, and Section VII concludes the paper. the next iteration of operation. The faded capacity is also
compared to the reference capacity fading defined by the
battery manufacturers in warranty terms (shown as capacity
II. BACKGROUND reference block in Fig. 1). If the estimated faded capacity
EMS managed storage devices are commonly deployed does not follow the reference standard trajectory, the health
by behind-the-meter (BTM) commercial and industrial (C&I) management unit limits the battery usage through its operation
customers for the two purposes of increasing renewable uti- parameters such as state of charge and the throughput (absolute
lization and reducing their demand charge, as depicted in accumulative value of battery usage). It will then pass the
Fig. 2. For the first case, storage device helps C&I customers updated limits to EMS to be implemented for the next iteration
to maximize the local consumption of their uncontrollable of operation.
renewable generation (mostly solar generations) to reduce
electricity bill charges. For the second use case, storage device III. R ELATED W ORK
helps to reduce the electricity bills by shaving peak electrical Capacity fading is a complex phenomenon of two factors,
demand, as explained in the following. namely cyclic aging and calendar aging [18]. Cyclic aging is
A C&I type electricity bill includes two main components, caused by charging and discharging actions over the course
namely energy and demand charges [14]. The energy compo- of the battery’s lifetime. It has been observed that the battery
nent shows the amount of energy (kilo Watt hour) consumed capacity is directly affected by charge/discharge characteristics
by the customer during a billing period that is typically a such as State of Charge (SoC), charge and discharge rates and
month. The demand charge (DC) is calculated based on the energy throughput [19]. Calendar aging is, however, caused
highest power demand (kilo Watt), which is measured at the when battery remains idle. Calendar aging is also a function
meter (point of connection to the power grid) during the same of battery age, ambient temperature and idle time SoC.
billing period. The demand charge is a significant portion of Several papers have studied the the process of battery cyclic
the bill (up to half of the total monthly charges [15]). aging [20]–[22]. In [20], the authors examined the cyclic aging
Battery device helps to minimize the DC by supplying the of battery cells based on the data collected from controlled
electrical demand during the peak hours [16], [17]. Fig. 3 experiments for which the operating conditions such as charge
shows an example daily demand and the battery operation. The and discharge rates were fixed. In [21], a particle filtering
figure shows that a storage solution is employed from 9:00 to framework considering the charge and discharge cycles is
14:00 when the customer experiences its peak demand. Using developed for modeling the cyclic aging of battery units.
the storage reduces the peak demand read at the meter and The authors of [22] also proposed an ensemble model to
consequently reduces the demand charge. After 14:00, battery characterize the capacity degradation and predict the remaining
is charged to be prepared for the next peak demand event. useful performance of lithium-ion batteries.
As a result, it is important to be able to accurately estimate There are also a number of research papers focusing on
the capacity in order to design an optimal battery charging and calendar aging [23]–[25]. In [23], a calendar life estimation
TABLE I: Battery Aging Dataset
Date/time Charge Power [kW] Discharge Power [kW] Temperature [C] SoC [%] Measured Capacity [kWh]
5/8/2013 7:04 PM 0.0 0.9 28 65 29.0
5/8/2013 7:09 PM 1.1 0.0 29 68 29.0
5/8/2013 7:14 PM 1.8 0.0 29 72 29.0
··· ··· ··· ··· ··· ···
was proposed based on five different datasets, where for each relatively few data points to train with. In order to im-
dataset, the ambient temperature is fixed but the charge/dis- prove the training performance, we propose to interpolate
charge rates are variable. The method investigated in [24] between the capacity measurements at times for which
considers the effect of temperature, SoC, and the idle time of battery operation features are available.
the battery on the calendar aging process. In [25], the authors In the following, we provide the details of the feature extrac-
developed a lifetime prediction approach based by quantifying tion and interpolation techniques.
the impact of temperature and SoC on impedance rise and
capacity loss.
In this paper, we develop a battery degradation model
A. Feature Extraction
based on the real data collected from battery storage solutions
installed and operated for behind-the-meter customers. Our We extract the features associated with cyclic and calendar
model is, thus, generic and does not rely on controlled operat- aging from the raw battery operation data. Since a BTM
ing conditions. Moreover, we integrate both cyclic and calen- battery installation usually experiences a complete operation
dar aging processes in a unified framework, by extracting the cycle in a weekly basis, we preprocess each week of operation
corresponding features from operation data. Our experimental separately. The cyclic aging features are defined as follows:
results confirm that, since batteries encounter both cyclic and w
• Crate : Weekly charge rate calculated based on the nor-
calendar aging on a daily basis, combining them can improve
malized charging rates as follows:
the prediction accuracy of the capacity degradation model.
P
t∈week Pcha (t)∆T (t)/C
w Pcha (t)6=0
IV. BATTERY O PERATION DATA Crate = P ,
t∈week ∆T (t)
Pcha (t)6=0
To model the battery degradation process, we collect real
time-series data obtained during three years of operation of
where Pcha (t) is the charging value at time t and with
different BTM battery installations. Table I shows sample
duration of ∆T (t), and C is the current battery capacity.
rows from the dataset. During operation, data of date and w w
• Drate : Weekly discharge rate calculated similar to Crate ,
time, battery charge power, battery discharge power, ambient
where Pcha (t) is replaced with Pdis (t), the discharging
temperature, and state of charge (SoC) are recorded every five
power at time t.
minutes. The battery capacity is, however, measured at about
• W hth : Weekly Watt hour throughput defined as absolute
every six months, and is shown as the same value until the next
accumulative value of battery usage, calculated by:
measurement is taken. Hence, to predict the battery capacity,
we can use six months of recorded operation information with X X
W hth = Pcha (t)∆T (t)+ Pdis (t)∆T (t).
five minute sampling time, which is equal to 259200 input data
t∈week t∈week
points. 1 Pcha (t)6=0 Pdis (t)6=0
When training a machine learning model on the battery
aging dataset, we face the following two challenges: • SoC ave : Average weekly SoC.
• ∆SoC: Difference between the initial and final SoC
• The battery operation data are recorded at a very small
measurements at the corresponding week.
time scale (five minutes). Such fine-grained measure-
ments are unlikely to have observable impact on the The calendar aging features are defined as follows:
battery capacity. Moreover, the large number of data • T : Battery age measured as the week number for which
samples causes both training and testing to be very we extract the weekly features.
slow. In order to reduce the input size and also preserve • Tidle : Weekly idle time is the weekly accumulated shelf
information, we propose to split the data into weekly time in which the battery is not charged or discharged.
blocks and extract one set of features for each block. • SoC idle : Idle SoC is the average SoC values during idle
• Unlike the battery operation data, there are very few times.
capacity measurements provided at a much larger time • T empave : Average weekly ambient temperature.
scale. This causes the machine learning model to have
Hence, the battery operation data of each week is summa-
1 Thetotal number of input data points in every six months is: 6 months × rized in nine features to be used as input to the machine learn-
30 days×24 hours×12 samples per hour×5 features per sample = 259200. ing model in order to predict the battery capacity degradation.
Actual Capacity Measurements 29.0 Actual Capacity Measurements
28 Interpolated Capacity Values Interpolated Capacity Values
28.5
Capacity (kW)
Capacity (kW)
26
28.0
24 27.5
22 27.0
20 26.5
0 50000 100000 150000 200000 250000 0 50000 100000 150000 200000 250000
Time Step (Five Minutes) Time Step (Five Minutes)
Fig. 4: Two examples of interpolated curves of capacity values.
interpolated values, i.e., it is 1 if and only 𝐶if𝑡 the corresponding Fig. 6: LSTM equations and a schematic of the LSTM unit. In
B. Our Method
capacity valueBattery is available or interpolated. Let w be a truncated We use equations,
time-series dataσ isof the
battery operations
sigmoid and ◦ denotes the entry-
and predict
function
Data 2m + 1Capacity ෪ battery capacity fading. The battery operations, such as charg-
Gaussian filter with points Δ𝐶 with variance𝐶ሚ𝑡+1 m/3. The
Operation Battery 𝑡 wise product. Variables h and c represent short- and long-term
ing or discharging rate, idle SoC time and temperature, can
iteration steps are statedFading Model
as follows: memories, respectively.
combine in a complex way to make long term impacts on
Fig.w1:~The Ck−1proposed battery capacity fading model predicts the the battery capacity. Therefore, an LSTM-based model is a
Ck ← change in capacity using, operation data. (1) tool to learn such variable-lag temporal dependencies.
suitable
w ~ Mk−1 The features extracted by the LSTM model are then fed to a
Ck (t) ← Ck−1 (t), if M0 (t) = 1, t = 1, 2, · · · , outputs the network
(2) fully-connected
one-layer capacityto makevalue. Our
the final proposed battery capacity
prediction.
Figure 5 fading
shows the model estimates
architecture used forthe change
capacity in capacity as shown in
fading
A. LSTM
Ck (t)←C k (t−1),Network
if Ck (t)>Ck (t−1) and Mk (t−1)=1,t=2,3,··· .prediction. (3) The LSTM unit and FC layer are set to have 50
Fig. 5. Hence, at each time step, the capacity can be predicted
• time-series nonlinear-¿RNN and 100 hidden neurons.
where ~•islong convolution
term -¿ LSTM
operation and division is element-wise. by adding the estimated change to the old capacity value.
At each iteration,
• Equations the curve is updated according to a 𝑦0 Recurrent neural𝑦𝑡−1network (RNN)
and figure
𝑦𝑡
is a class of artificial
normalized weighted
Recurrent neuralaverage
network filtering
(RNN) is usinga classonly the available neural networks with sequential units and hence is suitable
of artificial
or interpolated
neural networks values with (Equ. 1).units
sequential Original
and hence measurements
is suitable are FC for learning dynamic FC temporalFCbehaviors in time-series data.
for learning
then replaced dynamic
back intotemporal behaviorscurve
the updated in time-series
(Equ. data. 2). At the One particular class of RNNs is Long Short Term Memory
One particular class of RNNs is Long Short Term Memory
end, if(LSTM)
at a time step the capacity value has increased, it is LSTM
networks, which is specifically designed to capture …
(LSTM) networks, LSTM which is LSTM specifically designed to capture
replaced with the value of the previous time
long term dependencies and nonlinear relationships between step if that value long term dependencies and nonlinear relationships between
is available
inputs andor output.
already interpolated
An LSTM (Equ. 3).
network maintains The number of inputs
a time-varying 𝑥0 and output. An 𝑥𝑡−1LSTM network 𝑥𝑡 maintains a time-varying
state that can be interpreted as the memory of the network,
iterations is set as twice as the number of iterations it takes for state that can be
Fig. 3: interpreted
LSTM. as the memory of the network,
and is updated at each time step by incorporating the new
all values
inputtosample.
be nonzero. In experiments,
The time-varying memory enables the modelm to be
we choose and is updated at each time step by incorporating the new
equal to to the number
capture of data temporal
more complex points for months. toFigureIn 4figureinput
three compared
behaviors 5, each input sample xi , is a set of features
sample. The time-varying memory enables the model
extracted from one week of data, and yi are the change in
shows stationary
two examples models such as autoregressive
of interpolated models.
curves of The LSTM values.
capacity capacity into capture
corresponding more complex
week. Some temporal
of yi values behaviors compared to
are actual
model and equations are provided in Fig. 2.
stationary
ones and some modelsThe
are interpolated. such as autoregressive
one sample per one week models. The LSTM
V. P ROPOSED 𝑥𝑡 M ETHOD is a trade-off
modelbetweenandkeeping modelare
equations complexity
provided to be
in low
Fig. 6.
and also to preserve the information in data. The number of
After preprocessing, we have a time-series data containing Battery
time steps is about operation
150 samples, which features
helps the such
LSTM as charging or discharging
model
one set of battery 𝑐
operation features per week and the to capture rate,
corre- idle
the temporal SoC time
correlations and temperature
in data. behave in a complex way
𝑡−1 𝑐 𝑡
sponding capacity value. Therefore, F we need a model capable to make long term impacts on the battery capacity. Therefore,
R EFERENCES
of processing time-series ℎ 𝑡−1
data thatG gets weekly ℎ 𝑡
features and an LSTM-based model is a suitable tool to learn such variable-
[1] L. Liao and F. Köttig, “Review of hybrid prognostics approaches
for remaining useful life prediction of engineered systems, and an
application to battery life prediction,” IEEE Transactions on Reliability,
vol. 63, no. 1, pp. 191–207, 2014.
[2] S. M. Rezvanizaniani, Z. Liu, Y. Chen, and J. Lee, “Review and recent
𝑦𝑡 advances in battery health monitoring and prognostics technologies for
𝑦0 𝑦𝑡−1 𝑦𝑡 TABLE II: Comparison of different methods for prediction of
capacity fading.
FC FC FC Model Persistence Model AR Model Proposed Model
MAE (kW) 0.84 0.41 0.22
VII. C ONCLUSION
lag temporal dependencies. We feed the features extracted by An accurate prediction of capacity fading helps energy
the LSTM model to a one-layer fully-connected (FC) network management system to better monitor the behavior of batteries
to make the final prediction. Figure 7 shows the architecture and give real-time suggestions to improve the product lifetime.
used for capacity fading prediction. In this figure, each input It also provides a better baseline for manufacturers to decide
sample xi , is a set of features extracted from one week of on warranty rules for their products. In this paper, we trained
data, and yi is the change in capacity in corresponding week. a deep learning model on real battery operation data to predict
Some of yi values are actual ones and some are interpolated. capacity fading. We first processed operation data of each
The number of time steps is about 150 samples, which helps week separately to extract the features associated with cyclic
the LSTM model to capture the temporal correlations in data. and calendar aging. We then interpolated between capacity
measurements at times for which battery operation features are
VI. E XPERIMENTAL R ESULTS available. The preprocessed data were used to train a model
with LSTM units followed by a fully-connected network.
We randomly split the data into three sets of training data
The experimental results showed that our method accurately
(70%), validation data (15%) and test data (15%). Validation
predicts capacity fading and significantly outperforms baseline
data is used to set the hyperparameters and to choose the
models, including persistence and autoregressive (AR) models.
best performing model. Training features are normalized to
In our future works, we will develop a health management
have zero mean and unit variance. Validation and test features
framework to complete the battery health aware module.
are also normalized using the mean and standard deviation of
training data. The LSTM unit and FC layer are set to have 50
R EFERENCES
and 100 hidden neurons, respectively.
The model is trained to minimize the Mean Absolute Error [1] S. Koohi-Kamali, V. Tyagi, N. Rahim, N. Panwar, and H. Mokhlis,
“Emergence of energy storage technologies as the solution for reliable
(MAE) with respect to the interpolated capacity values. The operation of smart power systems: A review,” Renewable and Sustain-
final test error is computed based on actual capacity measure- able Energy Reviews, vol. 25, pp. 135–165, 2013.
ments. For this, we aggregate the deviations from interpolated [2] W. Zhang, M. Qiu, and X. Lai, “Application of energy storage technolo-
gies in power grids,” Power System Technology, vol. 7, 2008.
measurements in one sampling period (six month) and report it [3] C. Wang, W. Sun, T. Yi, Z. Yan, and Y. Zhang, “Review on energy
as the error in the corresponding time period. We then compute storage application planning and benefit evaluation methods in smart
the mean absolute values of all deviations as the final error. grid,” in Proceedings of the Chinese Society of Electrical Engineering,
vol. 33, pp. 33–41, 2013.
The model performance is compared with persistence and [4] A. Hooshmand, J. Mohammadpour, H. Malki, and H. Daneshi, “Power
autoregressive (AR) models. Persistence method assumes that system dynamic scheduling with high penetration of renewable sources,”
the next measurement of capacity will be the same as current in American Control Conference (ACC), pp. 5827–5832, IEEE, 2013.
[5] S. F. Tie and C. W. Tan, “A review of energy sources and energy
capacity, i.e., capacity does not degrade. AR model is a linear management system in electric vehicles,” Renewable and Sustainable
function of current and historical battery operation features. Energy Reviews, vol. 20, pp. 82–102, 2013.
Let X be the feature matrix of size N × d, where N and d [6] A. Khaligh and Z. Li, “Battery, ultracapacitor, fuel cell, and hybrid
energy storage systems for electric, hybrid electric, fuel cell, and plug-in
are number of samples and number of features, respectively. hybrid electric vehicles: State of the art,” IEEE transactions on Vehicular
In matrix X, each row contains features of one week. The AR
Pq−1 Technology, vol. 59, no. 6, pp. 2806–2814, 2010.
model of order q can be formulated as ỹt = τ =0 Xt−τ Aτ + [7] W. Kempton, V. Udo, K. Huber, K. Komara, S. Letendre, S. Baker,
D. Brunner, and N. Pearre, “A test of vehicle-to-grid (v2g) for energy
β, where Xi is the i-th row of X, i.e., features of time-step i, storage and frequency regulation in the pjm system,” Results from an
Ai is a vector of size d×1, and β is a scalar. In our dataset, we Industry-University Research Partnership, vol. 32, 2008.
have N = 150 and d = 9, and we use all available historical [8] D. Kottick, M. Blau, and D. Edelstein, “Battery energy storage for
frequency regulation in an island power system,” IEEE Transactions
data for training and testing the model. on Energy Conversion, vol. 8, no. 3, pp. 455–459, 1993.
Figure 8 shows the results of predicted capacity fading for [9] X. Liu, A. Aichhorn, L. Liu, and H. Li, “Coordinated control of dis-
three different batteries. As can be seen, our proposed model tributed energy storage system with tap changer transformers for voltage
rise mitigation under high photovoltaic penetration,” IEEE Transactions
accurately predicts weekly capacity degradation and provides a on Smart Grid, vol. 3, no. 2, pp. 897–906, 2012.
close estimate for the actual capacity. The AR model, however, [10] A. Hooshmand, Predictive Energy Management Methods for Smart
fails to track the capacity changes and thus produces results Grids. PhD thesis, 2012.
[11] R. Zamora and A. K. Srivastava, “Controls for microgrids with storage:
with large deviations from actual values. Table II provides the Review, challenges, and research needs,” Renewable and Sustainable
MAE results. The proposed model significantly outperforms Energy Reviews, vol. 14, no. 7, pp. 2009–2018, 2010.
0 0 29 Actual Capacity
Weekly Faded Capacity (W)
Capacity (kW)
40 40 27
60 60
26
80 80
25
100 Actual Faded Capacity 100 Actual Faded Capacity
Predicted Using AR Model Predicted Using Our Model 24
120 120
0 20 40 60 80 100 120 140 160 0 20 40 60 80 100 120 140 160 0 20 40 60 80 100 120 140 160
Time Step (Week) Time Step (Week) Time Step (Week)
(a)
0 0 Actual Capacity
Actual Faded Capacity 29
Weekly Faded Capacity (W)
Capacity (kW)
30 30 27
40 40
26
50 50
25
60 60
Actual Faded Capacity
70 70 Predicted Using Our Model 24
0 20 40 60 80 100 120 140 160 0 20 40 60 80 100 120 140 160 0 20 40 60 80 100 120 140 160
Time Step (Week) Time Step (Week) Time Step (Week)
(b)
0 0 29 Actual Capacity
Weekly Faded Capacity (W)
Capacity (kW)
20 20
27
30 30
40 40 26
50 50
25
60 Actual Faded Capacity 60 Actual Faded Capacity
70 Predicted Using AR Model 70 Predicted Using Our Model 24
0 20 40 60 80 100 120 140 160 0 20 40 60 80 100 120 140 160 0 20 40 60 80 100 120 140 160
Time Step (Week) Time Step (Week) Time Step (Week)
(c)
Fig. 8: Comparison of autoregressive (AR) and proposed methods for capacity fading prediction for three different batteries.
Left column compares the predicted faded capacity of the AR model with actual values. Middle column compares the predicted
faded capacity of our proposed model with actual values. Right column shows the predicted capacity using the AR and proposed
models along with the actual curve obtained by interpolating between capacity measurement (star points). As can be seen, our
proposed model accurately predicts the weekly capacity degradation and provides a close estimate for the actual capacity. The
AR model, however, fails to track the capacity changes and thus produces results with large deviations from actual values.
[12] A. Hooshmand, B. Asghari, and R. K. Sharma, “Experimental demon- [19] S. A. Pourmousavi, B. Asghari, and R. K. Sharma, “A novel algorithm to
stration of a tiered power management system for economic operation of integrate battery cyclic and calendar agings within a single framework,”
grid-tied microgrids,” IEEE Transactions on Sustainable Energy, vol. 5, in Innovative Smart Grid Technologies Conference (ISGT), Power &
no. 4, pp. 1319–1327, 2014. Energy Society, pp. 1–5, IEEE, 2016.
[13] A. Hooshmand, R. Sharma, and B. Asghari, “Multi-objective energy [20] F. Rufus, S. Lee, and A. Thakker, “Health monitoring algorithms for
management methods for micro-grids,” Feb. 14 2017. US Patent space application batteries,” in Prognostics and Health Management,
9,568,901. 2008. International Conference on, pp. 1–8, IEEE, 2008.
[14] “Electric rates,” 2018. PG&E Corporation, available at [21] B. Saha and K. Goebel, “Modeling li-ion battery capacity depletion in
https://www.pge.com/nots/rates/tariffs/electric.shtml. a particle filtering framework,” in Proceedings of the annual conference
of the prognostics and health management society, San Diego, 2009.
[15] A. Oudalov, R. Cherkaoui, and A. Beguin, “Sizing and optimal operation
[22] Y. Xing, E. W. Ma, K.-L. Tsui, and M. Pecht, “An ensemble model for
of battery energy storage system for peak shaving application,” in Power
predicting the remaining useful performance of lithium-ion batteries,”
Tech, 2007 IEEE Lausanne, pp. 621–625, IEEE, 2007.
Microelectronics Reliability, vol. 53, no. 6, pp. 811–820, 2013.
[16] R. Moslemi, A. Hooshmand, and R. K. Sharma, “A data-driven de- [23] Q. Zhang and R. E. White, “Calendar life study of li-ion pouch cells,”
mand charge management solution for behind-the-meter storage appli- Journal of Power Sources, vol. 173, no. 2, pp. 990–997, 2007.
cations,” in Power & Energy Society Innovative Smart Grid Technologies [24] J. Belt, V. Utgikar, and I. Bloom, “Calendar and phev cycle life aging
Conference (ISGT), pp. 1–5, IEEE, 2017. of high-energy, lithium-ion cells containing blended spinel and layered-
[17] R. Moslemi, A. Hooshmand, and R. K. Sharma, “A machine learning oxide cathodes,” Journal of Power Sources, vol. 196, no. 23, 2011.
based demand charge management solution,” in Innovative Smart Grid [25] M. Ecker, J. B. Gerschler, J. Vogel, S. Käbitz, F. Hust, P. Dechent, and
Technologies Conference Europe, IEEE PES, pp. 1–6, IEEE, 2017. D. U. Sauer, “Development of a lifetime prediction model for lithium-
[18] C. Schlasza, P. Ostertag, D. Chrenko, R. Kriesten, and D. Bouquain, ion batteries based on extended accelerated aging test data,” Journal of
“Review on the aging mechanisms in li-ion batteries for electric vehicles Power Sources, vol. 215, pp. 248–257, 2012.
based on the fmea method,” in Transportation Electrification Conference
and Expo (ITEC), 2014 IEEE, pp. 1–6, IEEE, 2014.