Remaining Useful Life Prediction for Lithium-

ion Battery Using Ensemble Learning Method

Bin Gou, Yan Xu, Sidun Fang, Ryan Arya Pratama Shuyong Liu
School of Electrical and Electronics Engineering Rolls-Royce Electrical
Nanyang Technological University Rolls-Royce @ NTU Corporate Laboratory
Singapore, Singapore Rolls-Royce Singapore
[email protected], [email protected], Singapore, Singapore
[email protected], [email protected] [email protected]

Abstract—Lithium-ion batteries (LIBs) have been widely model-based method and data-driven method. Model-based
applied in energy storage system, and its prognostics and health approaches formulate the degradation process of battery and
management are of great importance for the system performance. sequentially estimate the model parameters from the measured
In this paper, an ensemble data-driven method is proposed to
accurately predict the remaining useful life (RUL) of lithium-ion data using filter techniques, such as Particle Filter (PF). For
batteries. A strongly related feature extracted from the charging instance, Zhai et al. [3] proposed a Wiener process model to
voltage is selected to indicate the degradation trend of battery, execute the degradation-based RUL prediction of the battery.
according to the Pearson correlation analysis. A random learning He et al. [4] used the Dempster–Shafer theory and the
algorithm named Random Vector Functional Link (RVFL) Bayesian–Monte Carlo method to predict the RUL of LIBs.
network is applied to map the knowledge relationship between the The PF is an effective model-based method and widely
extracted feature and practical health state of the battery due to
its fast learning speed. The nonlinear autoregressive with applied to predict the RUL of systems. The core idea of the PF
exogenous inputs (NARX) structure which contains past and is to use random particles with their associated weights to
present information is introduced to set up the deterioration approximate the posterior probability distributions of states
process of the LIBs. An ensemble learning method that combines through sequential importance sampling. Saha et al. [5]
a set of single RVFL models is designed to further improve the proposed a PF-based prognostics method for RUL prediction of
RUL prediction accuracy. The test results using publicly available battery. Dong et al. [6] investigated long-term RUL prediction
dataset show that the proposed ensemble data-driven method can
accurately predict the RUL of batteries without any other using Brownian motion based degradation model and PF. Yu et
additional hardware or system downtime. al. [7] developed a battery health prognostics system based on
Bayesian-inference probabilistic indication and integrated
Index Terms— Lithium-ion batteries (LIBs), remaining useful logistic regression and PF, where PF model is applied to predict
life (RUL), Random Vector Functional Link (RVFL) network, the dynamic degradation propagation of the battery health. In
ensemble learning. [8], A particle learning framework integrated with smoothing
algorithm is proposed to predict the RUL of LIBs.
I. INTRODUCTION
In summary, model-based methods have obtained success in

R ECENTLY, the Energy Storage System (ESS) is widely

applied into microgrid system and electric vehicle since it
can achieve enormous benefits on both the economy or the
battery RUL prediction. However, the model-based methods
suffer from heavy computational burden, such as PF and Monte
Carlo method. Moreover, the accurate analytical model of the
environmental aspects, as well as expand the operating complex dynamic system is usually difficult to derive,
flexibility. Lithium-ion batteries (LIBs) as a high energy especially when the system operates under highly fluctuating
density storage system, become a core component of ESS due environments.
to the low self-discharge rate, less maintenance requirement, Data-driven methods have received extensive attention from
and long lifetime [1]. But as we all know, the performance of academia and industry due to their flexibility and model-free
the ESS would gradually deteriorate over time because of the characteristics. In these methods, machine learning algorithms
aging process [2]. are used to construct the health degradation process and predict
The unexpected failure of Hybrid Energy Storage System the RUL of battery from the historical database. Liu et al. [9]
(HESS) may result in system shutdown, adopted a data-driven monotonic echo state networks algorithm
unscheduled maintenance, or even catastrophic consequences. to track the nonlinear patterns of battery degradation. This
The predictive health condition analysis can help to prevent the approach constitutes a probabilistic integration and data-driven
unexpected breakdowns and optimally design the maintenance prognostic framework with uncertainty management capability.
schedule and inventory for stocks and spares, thereby reducing Based on the degradation data of LIBs from NASA, Liu et al.
operation and maintenance cost. [10] proposed an adaptive recurrent neural network for system
For LIBs, Remaining Useful Life (RUL) is applied to denote dynamic state forecasting for RUL prediction. In [11], the
the period from the present time to the end of the battery useful sample entropy of short voltage sequence is used as an effective
life. In recent years, a variety of approaches have been reported signature of health. Advanced sparse Bayesian predictive
for RUL prediction of LIB and can be generally categorized into modeling methodology is introduced to capture the underlying

978-1-7281-1981-6/19/$31.00 ©2019 IEEEI978-1-7281-1981-6/19/$31.00 ©2019 IEEEE978-1-7281-1981-6/19/$31.00 ©2019 IEEEE978-

Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on June 15,2020 at 00:04:15 UTC from IEEE Xplore. Restrictions apply.
correspondence between the battery health and sample entropy. I. Offline Ensemble learning Model Training
Data-driven approaches can achieve health state assessment Charging Input Training Ensemble
Historical Health Indicators RVFL
and reliability estimation based on testing data samples and Battery Cells Voltage Extraction & Evaluation Learning Method Model
monitoring parameters. But traditional data-driven approaches
always suffer from heavy computational burden and slow II. Online RUL Prediction

training speed. Operational

Charging
Online Health Input Output RUL
Ensemble Model
Recently, an emerging random learning technology named Battery Cell Voltage Indicators Extraction Prediction

Random Vector Functional Link (RVFL) network is applied in

Fig. 1. The RUL prediction structure based on RVFL ensemble learning.
the industrial applications due to their much faster learning
speed, better generalization capacity, and computationally
Input Data Learning Method Predicted Result
efficient tuning mechanism. The learning speed of RVFL can
Capacity(t) x1
be thousands of times faster than traditional learning algorithms i1
(such as support vector machine, SVM) and requires much less Z -1
h1
computation memory. Capacity(t-1) x2
i2
Therefore, in this paper, an ensemble RVFL learning Z -1
algorithm is applied to accurately predict the RUL of the LIBs.

...

...
...
Based on the quantitative correlation analysis, a feature with
Capacity(t-d) xd
strong correlation to the capacity of battery is extracted as the id h2 Capacity(t+p)
o1
health indicator to predict the future capacity. RVFL DECVD(t) xd+1
id+1
technology is applied to mapping the knowledge relationship
Z -1
between the selected health indicator and practical capacity.

...
DECVD(t-1) xd+2
The nonlinear autoregressive with exogenous inputs (NARX) id+2
structure which contains past and present information is Z -1 hK
introduced to set up the deterioration process of the LIBs.

...
...
...

Finally, an ensemble learning method is designed to improve DECVD(t-d) x2d

i2d
the RUL prediction accuracy. Fig. 2. The prediction structure of NARX.

II. PROPOSED METHODOLOGY Moreover, the ensemble model is designed to further improve
the prediction accuracy by combining a set of individual
The health of LIBs is generally reflected on parameters as
RVFLs.
total useful capacity reduction and resistance increase. In this
In the section of online RUL prediction, the health indicator
paper, capacity is used to indicate the health state of battery. In
of battery in operation is extracted from the online
general, a LIB is deemed to reach its end of life when its
measurements, and then input to the well-trained RUL
capacity fades by 20% or 30% of the rated value. When it
ensemble learning model, which is obtained in the offline stage.
reaches the threshold, the battery is considered as an unreliable
power source and should be replaced in time. The RUL is the B. Health Indicator Extraction and Evaluation
timespan from the present time to the defined end of life. In [12], the time duration of equal terminal voltage interval
A. Structure of the Framework during the discharging process is used as a health indicator to
measure the capacity degradation. But generally, the
In this paper, an ensemble learning method is proposed to
discharging process highly depends on the load profiles.
predict the RUL of LIB, and its structure is illustrated as Fig. 1.
Different discharging currents result in different discharging
The method comprises two stages: offline ensemble learning
voltage difference during the same time interval, affirmatively.
model training and online RUL prediction.
In this paper, the duration of equal charging voltage difference
In the section of offline training, the historical operational
(DECVD) is extracted to indicate the battery capacity
data such as voltage, current, and temperature are collected to
degradation since the charging process normally can be
set up a database. The practical capacities are calculated or
controlled by the battery management system, such as electric
estimated as the benchmark values. Based on the quantitative
vehicle applications. The DECVD is defined as
correlation analysis, a feature which has a strong relationship
with the practical capacity is adopted as the health indicator. A DECVD(Vmin ,Vmax ) = tVmax − tVmin (1)
fast and random learning technologies called RVFL network where tVmax and tVmax are the elapsed time when the terminal
are applied to extract the relationship between the selected voltage of battery reaches Vmax and Vmin, respectively. Vmax and
health indicator and the practical capacity and to predict the Vmin are the maximum and minimum of the selected terminal
health deterioration trend of the battery. NARX structure is voltage interval during the charging process.
introduced since it can contain the present and past information, Moreover, the Pearson correlation analysis is applied to
which better reflects the aging trend prediction of the battery. quantitatively evaluate the relationship between different
The RUL prediction structure of NARX is illustrated as Fig. 2. DECVD values and the practical capacity to select the most
The prediction principle is by integrating the present and past correlative health indicator. The Pearson correlation coefficient
health indicators and capacities to predict the future capacity. is computed as

Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on June 15,2020 at 00:04:15 UTC from IEEE Xplore. Restrictions apply.
 n
 i =1 ( xi − x )( yi − y )
Input Hidden Output
Layer Layer Layer
r= (2)
n n
 i =1 ( xi − x )2  i =1 ( yi − y )2
h1
i1
where r is between -1 and +1, and a perfect linear relationship h2 o1

exists between the independent variable x and observation y i2

when the correlation coefficient r = ± 1. r = 0 indicates no linear

...

...
i3
relationship between x and y. In this paper, x is the DECVD and
y is the practical capacity.

...
od
iL
III. ENSEMBLE LEARNING METHOD hK

As already introduced, traditional data-driven algorithm Fig. 3. The structure of RVFL learning algorithm.
always suffers from heavy training burden. In this paper, RVFL diversity, a set of individual learners can compensate for each
is applied for ensemble learning since RVFL is a randomized other and make a plurality decision. Therefore, the ensemble
learning algorithm, which has fast learning speed and learning structure could reduce the aggregated variance and
computationally efficient tuning mechanism. improve the accuracy and the generalization capacity over the
A. Random Vector Functional Link Neural Network individuals.
The proposed RVFL ensemble learning process can be
The key of the RVFL algorithm is to randomly choose the
described as follows:
input weights and biases, and analytically determine the output
weights through simple matrix computation. There are direct RVFL Ensemble Learning Rule
links between the input and output layers, i.e., the red arrows Given E single RVFLs and a database of F × Q size (where F and Q are
illustrated in Fig. 3, which is the biggest difference with other the total number of features and instances, respectively),
for i=1 to E:
learning technologies. A training set is denoted by the input data
Randomly sample q instances out of the database, 1  q  Q .
samples as X and the corresponding output target as Y. The
Randomly assign the i-th RVFL h hidden nodes and an activation
input features are first linearly transformed by the weights A function, h is within the optimal range [hmin, hmax], which is subject to a pre-
between the input and hidden layer, where the weights A are tuning procedure).
randomly generated. After that, a nonlinear activation function Train this RVFL and return.
End
h is applied to the transformed features to get the features in the
hidden layer. The features in the enhancement nodes can be C. NARX Structure
denoted as The battery health conditions are time-series and dynamic
H = h( XA ) . (3) systems. In this paper, the NARX structure is introduced to
Both features H and the input features X will be further sent improve the RUL prediction accuracy. Since the structure of
forward to the output layer. This configuration maps both linear NARX has feedback connections, it contains the present and
and nonlinear relationships. Consequently, it helps to regularize past information and can build autoregressive models, i.e.,
the enhancement features and improves the performance of models whose current output depends on past values of the
network[13]. output itself.
For RVFL, only the output weights Φ need to be optimized. In a conventional static structure, there is no exogenous
Suppose that the input data have L features and there are K input. The prediction network can be represented by
enhancement neurons, in total L + K inputs to each output node. yˆ( xi ) = f ( xi ) . (5)
Learning is achieved by minimizing the following expression When NARX structure is applied into the prediction process,
when mean square loss function is employed: the network can be rewritten as
1 N yˆ(t +p) = f (x(t ), x(t − 1), , x(t − d ),
 Yi − i Φ
2
= (4) (6)
2 N i =1 2 y (t − 1), y (t − 2), , y (t − d ))
where ℏ denotes the concatenation of the input features X and where x(t-d) denotes the previous health indicator of d time
enhancement features H, Φ is the output weights. step, y(t-d) is the previous corresponding capacity of d time
step, yˆ(t + p) denotes the predicted future capacity of p step. As
B. Ensemble RVFL Learning Network
illustrated in Fig.2, the structure provides a prediction depended
As already introduced, RVFL adopts random input weights on the past and present values of the health indicator and
for learning. Thus, its learning speed is faster than conventional capacity. In Fig. 2, z-1 denotes a time delay of one sample. The
ANNs. But this randomness could result in unstable outputs of future p cycles capacity can be predicted as long as the present
RVFL in practical applications due to the undesired side effects. and past d cycles health indicator (DECVD) and the
To overcome this issue, a series of single RVFLs are assembled corresponding capacity are inputted. Then, the RUL of the
in our proposal. Each of the individual RVFL chooses random battery could be calculated according to the predicted capacity.
input weights and other training parameters, including training
instances, hidden nodes, and activation function. Due to the

Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on June 15,2020 at 00:04:15 UTC from IEEE Xplore. Restrictions apply.
 1.2 0.05
Capacity of CS35
Capacity of CS36 0.045 sigmoid
1 Capacity of CS37 hmin hmax sine
Capacity of CS38 0.04 tribas
radbas

Validation test errors

0.8 0.035
Capacity (Ahr)

0.03
0.6
0.025
0.4
0.02

0.2 0.015

0.01
0 0 100 200 300 400 500 600 700 800 900 1000
0 200 400 600 800 1000 1200 Hidden nodes numbers
Cycles
Fig. 4. The attenuation curves of the capacity with cycle numbers. Fig. 5. Validation test for RVFLs with different numbers of hidden neurons.
0.05

0.045 With the past capacities of battery

IV. RUL PREDICTION RESULTS Without the past capacities of battery
0.04
In this paper, a widely used public battery aging test dataset

Validation test errors

0.035
from Center for Advanced Life Cycle Engineering (CALCE) of
0.03
the University of Maryland is introduced for validating the
0.025
effectiveness of the proposed ensemble data-driven method.
0.02
The cyclic test was implemented with the Arbin BT2000
Battery Test System under room temperature. In this paper, four 0.015

batteries named CS35, CS36, CS37, and CS38 are chosen as 0.01
0 100 200 300 400 500 600 700 800 900 1000
benchmark data given to the same experimental condition and Hidden nodes numbers

rated capacity (1.1 Ah). Fig. 4 shows the attenuation curves of Fig. 6. Validation test results with the historical capacity degradation
information.
the capacity with cycle numbers of these four cells. The end-of-
life (EOL) threshold is recommended to be around 80% of the prediction accuracy. As shown in Fig. 2, only the past and
rated capacity. future cycle number d and p need to be properly selected. The
prediction cycle number p is the objective which is set as 50 in
A. Parameters Tuning this paper. The future 50 cycles capacity can be predicted once
According to Eq. (1) and Eq. (2), the most correlated feature the present and past DECVDs and capacities are inputted to the
can be calculated. The Vmax and Vmin of DECVD are 3.8V and well-trained ensemble learning model. The past cycle number
4.2V, respectively. For RVFL, only the number of hidden layer d is selected according to the future cycle number p, generally
neurons and activation function need to be tuned properly. In bigger than p, i.e., 60 in this paper.
this paper, root mean square error (RMSE) is used to assess the Fig. 6 shows the RMSE versus single RVFL hidden nodes
training accuracy of single RVFL, defined as validation result with the historical capacity degradation
1 N Train information. It is clear that the training accuracy is higher when
RMSE _Train =
N Train
 (ti − g (x i ))2 (7) the past capacities are contained in the input features of the
i =1 learning model. Thus, in this paper, the past capacities of
where Ntrain is the number of instances in training dataset, ti is battery are also as one kind of the input features.
the target of each instance and g(xi) is the prediction result of
the single RVFL. The lower RMSE means the higher training B. Prediction Results and Analysis
accuracy. In this test, given the datasets of the four batteries, the data
As introduced in Section III-B, [hmin, hmax] is the range of of three batteries are integrated to predict the RUL of the
optimal hidden nodes for a single RVFL which is selected at the remaining battery. For instance, to predict the RUL of battery
highest validation accuracy range. For activation function, four CS 38, the data of CS 35, CS 36, and CS 37 are assembled as
initial functions are considered, i.e., ‘sigmoidal’, ‘sine’, training sets. In this paper, the absolute error (AE) is used to
‘triangular basis’ and ‘radial basis’, where the highest training qualify the error of the prediction results, shown as
accuracy by the function is selected as the best tuning of RVFL AE= RUL real − RUL predict . (8)
setting parameters. In this paper, data from the three batteries
The RUL prediction results of CALCE data are illustrated in
are integrated as the training data, and the remaining battery
Fig. 7. Since the parameters of d and p are selected as 60 and
data is selected as the test data. Fig. 5 shows the RMSE versus
50, respectively, the capacities of the 111th cycle to 160th cycle
single RVFL hidden nodes validation result. It can be seen that
can be predicted according to the structure illustrated in Fig. 2.
RVFLs will have a stable performance once the hidden nodes
When the health indicators (DECVDs) of the 111th cycle to
exceed 400. Thus, [hmin, hmax] can accordingly be set as [400,
160th cycle are measured, the capacities of the 161 st cycle to
600] with the activation function of ‘sine’ to ensure high
210th cycle can be predicted periodically. Thus, the capacity is
efficiency and low RMSE, simultaneously. For the ensemble
always predicted 50-cycle-ahead. Taking battery CS 35 as an
learning, the total learning model number is set as 200.
example, the predicted EOL is 608th cycle while the
In this paper, NARX structure is introduced to improve the
experimental EOL is 603rd cycle. The predicted absolute error

Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on June 15,2020 at 00:04:15 UTC from IEEE Xplore. Restrictions apply.
 1.2 1.2
Experimental Capacity Experimental Capacity
1 Predicted Capacity 1 Predicted Capacity
EOL Threshold EOL Threshold
Capacity (Ahr)

Capacity (Ahr)
0.8 0.8

0.6 Predicted EOL 0.6

Predicted EOL
0.4
Experimental
++ 0.4
++
EOL Experimental
0.2 0.2 EOL
603 Cycles 608 Cycles 548 Cycles 552 Cycles
0 0
100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900 1000
Cycles Cycles

(a) (b)
1.2 1.2
Experimental Capacity Experimental Capacity
1 Predicted Capacity 1 Predicted Capacity
EOL Threshold EOL Threshold
Capacity (Ahr)

Capacity (Ahr)
0.8 0.8

0.6 0.6
Predicted EOL Experimental
EOL
0.4
++ 0.4 Predicted EOL

0.2
Experimental
EOL 0.2 ++
618 Cycles 621 Cycles 649 Cycles 653 Cycles
0 0
200 400 600 800 1000 1200 200 400 600 800 1000 1200
Cycles Cycles

(c) (d)
Fig. 7. The RUL prediction result of batteries, (a) CS 35, (b) CS 36, (c) CS 37, (d) CS 38.
TABLE I Carlo method,” J. Power Sources, vol. 196, no. 23, pp. 10314–10321, Dec.
RUL PREDICTION RESULTS 2011.
Actual Predicted [2] W. Jing, C. H. Lai, W. S. H. Wong, and M. L. D. Wong, “A
Batteries RMSE AE
EOL Value EOL Value comprehensive study of battery-supercapacitor hybrid energy storage
CS 35 603 608 0.0243 5 system for standalone PV power system in rural electrification,” Applied
CS 36 538 552 0.0242 4 Energy, vol. 224, pp. 340-356, 2018.
CS 37 618 621 0.0255 3 [3] Q. Zhai and Z. Ye, “RUL prediction of deteriorating products using an
CS 38 649 653 0.0184 4 adaptive wiener process model,” IEEE Trans. Ind. Informat., vol. 13, no.
6, pp. 2911–Dec, Jul. 2017.
is as low as 5 cycles, which is acceptable in the electrical [4] W. He, N. Williard, M. Osterman, and M. Pecht, “Prognostics of lithium-
ion batteries based on Dempster–Shafer theory and the Bayesian Monte
applications.
Carlo method,” J. Power Sources, vol. 196, no. 23, pp. 10314–10321,
According to Fig. 7, the prediction results are summarized in Dec. 2011.
Table I. The predicted capacities can accurately trace the [5] B. Saha, K. Goebel, S. Poll, and J. Christophersen, “Prognostics methods
experimental capacities, which are measured during the aging for battery health monitoring using a Bayesian framework,” IEEE Trans.
Instrum. Meas., vol. 58, no. 2, pp. 291–296, Feb. 2009.
test. The RMSE prediction errors are within 0.0255 and RUL [6] G. Dong, Z. Chen, J. Wei, and Q. Liang, “Battery health prognosis using
prediction errors are within 5 cycles. Therefore, the proposed Brownian motion modeling and particle filtering,” IEEE Trans. Ind.
method can accurately predict the RUL of the LIBs. Electron., to be published.
[7] J. Yu, “State-of-health monitoring and prediction of lithium-ion battery
using probabilistic indication and state-space model,” IEEE Trans.
V. CONCLUSIONS Instrum. Meas., vol. 64, no. 11, pp. 2937–2949, Nov. 2015.
[8] Z. Liu, G. Sun, S. Bu, J. Han, X. Tang, and M. Pecht, “Particle learning
A new data-driven method based on ensemble RVFL for framework for estimating the remaining useful life of lithium-ion
battery RUL prediction is proposed in this paper. A feature, batteries,” IEEE Trans. Instrum. Meas., vol. 66, no. 2, pp. 280–293, Feb.
DECVD, which is extracted from the charging voltage signal, 2017.
[9] D. Liu, W Xie, H. Liao, and Y. Peng, “An integrated probabilistic
is selected as the health indicator. RVFL is applied as the approach to lithium-ion battery remaining useful life estimation,” IEEE
predictor to obtain the relationship between DECVD and Trans. Instrum. Meas., vol. 64, no. 3, pp. 660–670, Mar. 2015.
capacity and predict the battery degradation trend. RVFL [10] J. Liu, A. Saxena, K. Goebel, B. Saha, and W. Wang, “An adaptive
recurrent neural network for remaining useful life prediction of lithium-
ensemble learning with NARX structure is designed to improve ion batteries,” in Proc. of Annual Conference of the Prognostics Health
the accuracy and stability of the prediction results. Finally, a Management Society, 2010, pp. 1–9.
standard data from CALCE is introduced to verify the [11] X. Hu, J. Jiang, D. Cao, and B. Egardt, “Battery health prognosis for
effectiveness of the proposed ensemble data-driven method. electric vehicles using sample entropy and sparse Bayesian predictive
modeling,” IEEE Trans. Ind. Electron., vol. 63, no. 4, pp. 2645–2656,
The experimental results show that the proposed method can Apr. 2016.
accurately and steadily predict the RUL with less than 5 cycles [12] D. Liu, J. Zhou, H. Liao, Y. Peng, and X. Peng, “A health indicator
error as long as the charging voltage is measured, which is quite extraction and optimization framework for lithium-ion battery
degradation modeling and prognostics,” IEEE Trans. Syst. Man Cybern.
suitable for the online practical application. Syst., vol. 45, pp. 915–928, Jun. 2015.
[13] L. Zhang and P. Suganthan, “A survey of randomized algorithms for
REFERENCES training neural networks,” Inf. Sci., vol. 364–365, pp. 146–155, 2016.

[1] W. He, N. Williard, M. Osterman, and M. Pecht, “Prognostics of lithium-

ion batteries based on Dempster–Shafer theory and the Bayesian Monte

Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on June 15,2020 at 00:04:15 UTC from IEEE Xplore. Restrictions apply.