Machine learning systems based on xgBoost and MLP neural network applied in satellite lithium-ion battery sets impedance estimation

Advanced Computational Intelligence: An International Journal, Vol.5, No.1, January 2018
Machine learning systems based on xgBoost
and MLP neural network applied in satellite
lithium-ion battery sets impedance estimation
Thiago H. R. Donato and Marcos G. Quiles
Department of Applied Computation, National Space Research Institute,
Sao Jose dos Campos, P.0 Box 12227-010, Brazil
Institute of Science and Technology, Federal University of Sao Paulo,
Sao Jose dos Campos, P.0 Box 12231-280, Brazil
Abstract. In this work, the internal impedance of the lithium-ion battery pack (important mea-
sure of the degradation level of the batteries) is estimated by means of machine learning systems
based on supervised learning techniques MLP - Multi Layer Perceptron - neural network and xg-
Boost - Gradient Tree Boosting. Therefore, characteristics of the electric power system, in which
the battery pack is inserted, are extracted and used in the construction of supervised models
through the application of two different techniques based on Gradient Tree Boosting and Multi
Layer Perceptron neural network. Finally, with the application of statistical validation techniques,
the accuracy of both models are calculated and used for the comparison between them and the
feasibility analysis regarding the use of such models in real systems.
Keywords: Lithium-ion battery, Internal impedance, State of charge, Multi Layer Perceptron,
Gradient Tree Boosting, xgBoost
1 Introduction
The choice for the technology to be applied in the electrical power system (EPS) is
important to the success of a satellite mission since it represents around 20 to 30
percent of a satellite total mass.
For outer space applications, lithium-ion batteries have less than one half of the
mass of nickel hydrogen batteries for the same stored energy [6] and, for this reason,
can reduce the system weight of a spacecraft, thus improving the load efficiency of
satellites. As a result, lithium-ion batteries are highly adopted in the satellites of
the United States and European Space Agency (ESA) [17]. In addition, lithium-ion
batteries are expected to become the third generation of satellite power storage
batteries for Chinas future space platform instead of NiMH batteries and NiCd
batteries.
The reliability of Li-ion batteries is an important issues due to the fact that
failures of Li-ion battery not only result in serious inconvenience and enormous re-
placement/repair costs, but also can cause overheating and short circuiting which
DOI: 10.5121/acii.2018.5101 1

can lead to catastrophic consequences, such as explosion. In order to prevent se-
vere failures from occurring, and to optimize Li-ion battery maintenance schedules,
breakthroughs in prognostics and health monitoring of Li-ion batteries, with an
emphasis on fault detection, correction and remaining-useful-life prediction, must
be achieved [20].
Battery PHM has a wide variety of meaning, ranging from irregular manual
measurements of voltage and electrolyte specific characteristics to fully automated
online observation of various measured and estimated battery parameters. An im-
portant battery PHM analysis consists of the measurement of its state of charge
(SOC). The state of charge (SOC) of Li-ion battery sets is a function of its inter-
nal impedance. Therefore, the state of charge (SOC) can be estimated through the
methods described on Section 2 or determined by the battery internal impedance
measurement or estimation along the operating cycles, as proposed in this study.
2 Li-ion battery state of charge (SOC) estimation
State of charge estimation has always been a big concern for all battery driven
devices but its definition presents many different issues [3]. In general, the SOC
of a battery is defined as the ratio of its current capacity (Qt) to the nominal
capacity (Qn). The nominal capacity is given by the manufacturer and represents
the maximum amount of charge that can be stored in the battery. The SOC can
be defined as follows:
SOCt =
Qt
Qn
. (1)
Therefore, once the nominal capacity is already defined, the methods described
below intend to estimate the actual battery capacity after charge/discharge cycles.
The various mathematical methods of estimation are classified according to
methodology. The classification of these SOC estimation methods is different in the
various literatures. One approach is according to the following categories [18]:
2.1 Direct measurements
Direct measurement methods refer to some physical battery properties such as
the terminal voltage and impedance. Many different direct methods have been
employed: open circuit voltage method, terminal voltage method, impedance mea-
surement method, and impedance spectroscopy method.
Open circuit voltage method There is approximately a linear relationship be-
tween the SOC of the lead-acid battery and its open circuit voltage (OCV ) given
by:
2

OCV (t) = a1 ∗ SOC(t) + a0. (2)
Where SOC(t) is the SOC of the battery at t, a0 is the battery terminal voltage
when SOC = 0, and a1 is obtained from knowing the value of a0 and OCV (t) at
SOC = 1. Unlike the lead-acid battery, the Li-ion battery does not have a linear
relationship between the OCV and SOC [4].
Terminal voltage method The terminal voltage method is based on the terminal
voltage drops because of the internal impedances when the battery is discharging,
so the electromotive force (EMF) of battery is proportional to the terminal voltage.
Since the EMF of battery is approximately linear proportional to the SOC, the
terminal voltage of battery is also approximately linear proportional to the SOC
[15]. The terminal voltage method has been employed at different discharge cur-
rents and temperatures but, at the end of battery discharge, the estimated error of
terminal voltage method is large, because the terminal voltage of battery suddenly
drops at the end of discharge.
Once the Li-ion battery does not have a linear relationship between the VOC
and SOC, this method is not indicated due to the estimation error.
Impedance spectroscopy method The impedance spectroscopy method mea-
sures battery impedances over a wide range of ac frequencies at different charge and
discharge currents. The values of the model impedances are found by least-squares
fitting to measured impedance values. SOC may be indirectly inferred by measur-
ing present battery impedances and correlating them with known impedances at
various SOC levels [10].
Coulomb counting method The Coulomb counting method measures the dis-
charging current of a battery and integrates the discharging current over time in
order to estimate SOC. Coulomb counting method is done to estimate the SOC(t)
which is estimated from the discharging current, I(t) and previously estimated
SOC values, SOC(t − 1). SOC is calculated by the following equation:
SOC(t) = SOC(t − 1) +
I(t)
Qn
∗ ∆t. (3)
2.2 Machine learning systems
Machine learning system consists of an approach which uses pattern recognition
and machine learning techniques to detect changes in system states [12]. With this
approach, few information regarding to the analyzed system is necessary in order
to build prognostic models due to the fact that only the monitored data itself is
3

necessary to elaborate them. For this reason, machine learning systems are ad-
equate to EPS systems which are sufficiently complex such that developing an
accurate physical model is prohibitively expensive. As batteries have been affected
by many chemical factors and have nonlinear SOC, machine learning systems offer
good solution for SOC estimation. However, this approach implies in wider confi-
dence intervals than other approaches and requires a substantial amount of data
for training.
After the data preparation, a machine learning algorithm shall be selected in
order to build the machine learning system. The selection of a proper algorithm
for a specific application is a challenging factor in applying data driven prognostics
methods. Examples of machine learning algorithms applied in machine learning
systems to estimate SOC: back propagation Multi Layer Perceptron neural network
(MLP), gradient boosting (xgBoost), radial basis function (RBF), fuzzy logic
methods, support vector machine (SV M), fuzzy neural network, and Kalman filter
[16].
In this study, two machine learning systems were applied: MLP neural network
(Section 2.2) and gradient boosting (Section 2.3)
Multi Layer Perceptron In order to build an algorithm capable of classifying a
label attribute, the Multi Layer Perceptron (MLP) neural network can be applied.
The MLP consists of a feedforward artificial neural network model that can be
used in classification or prognostic issues.
The Multi Layer Perceptron network contains three or more layers (an input and
an output layer with one or more hidden layers) of nonlinearly-activating neurons.
Each neuron combine the inputs multiplied by their correspondent weights and
apply an activation function which output is delivered as input of neurons of the
following layer [8, Chapter 4].
In many practical applications of artificial neural networks (ANN), there ex-
ist natural constraints on the model such as monotonic relations between inputs
and outputs that are known in advance. It is advantageous to incorporate these
constraints into the ANN structure [19]. The monotonic Multi Layer Perceptron
network (MONMLP) is an approach for multi-dimensional function approximation
ensuring monotonicity for selected input-output relations. Moreover, we determine
the requirements for the network structure regarding universal approximation ca-
pabilities.
2.3 Gradient Tree Boosting
Gradient tree boosting is typically used with decision trees (especially CART trees)
of a fixed size as base learners. For this special case Friedman proposes a modifi-
cation to gradient boosting method which improves the quality of fit of each base
learner.
4

Generic gradient boosting at the m − th step would fit a decision tree hm(x)
to pseudo-residuals. Let Jm be the number of its leaves. The tree partitions the
input space into Jm disjoint regions R1m, . . . , RJmm and predicts a constant value
in each region. Using the indicator notation, the output of hm(x) for input x can
be written as the sum:
hm(x) =
Jm
X
j=1
bjmI(x ∈ Rjm), (4)
where bjm is the value predicted in the region Rjm.
Then the coefficients bjm are multiplied by some value γm, chosen using line
search so as to minimize the loss function, and the model is updated as follows:
Fm(x) = Fm−1(x)+γmhm(x), γm = γarg min
n
X
i=1
L(yi, Fm−1(xi)+γhm(xi)) (5)
Friedman proposes to modify this algorithm so that it chooses a separate optimal
value γjm for each of the tree’s regions, instead of a single γm for the whole tree.
He calls the modified algorithm ”TreeBoost” [7]. The coefficients bjm from the tree-
fitting procedure can be then simply discarded and the model update rule becomes:
Fm(x) = Fm−1(x)+
Jm
X
j=1
γjmI(x ∈ Rjm), γjm = γarg min
X
xi∈Rjm
L(yi, Fm−1(xi)+γ)
(6)
Root Mean Square Error Machine learning systems apply machine learning
techniques in a supervised approach. Considering a numeric label attribute in the
estimative of the battery set impedance, each observed value can be compared with
the predicted one. This individual deviation is called a residual and the aggregation
of all the residuals is denominated the Root Mean Square Error (RMSE), obtained
as follows [9]:
RMSE =
sPn
t=1(ŷt − yt)2
n
(7)
Where:
– ŷt: predicted instance
– yt: observed instance
– n: number of instances
5

2.4 Li-ion battery state of charge (SOC) based on battery internal
impedance
According to the definition of the SOC, the function which provides the battery
state of charge based on battery internal impedance is obtained as follows:
SOCt = SOCt0 +
Z t
t0
(
η · I
Ct
) · dt (8)
Where:
– SOCt0: estimated SOC at time t0, when the estimation process starts
– SOCt: estimated SOC at time t
– η: current efficiency
– I: current - assumed to be positive when charging
– Cn: capacity of the battery at time t
In this study, the battery impedance is obtained through the machine learning
systems (see Section 2.2) which can be applied in the determination of battery state
of charge (SOC).
Bagging optimization According to bagging optimization method, a training
set Dofsizen is divided into m new training sets Di, eachofsizen, by sampling
from D uniformly and with replacement. Sampling the m new training sets with
replacement, implies that some observations may be repeated in each Di[1].
This kind of sample is also known as bootstrap sample. In order to obtain the
resulting ensemble of models, the m models are fitted using the above m bootstrap
samples. Finally, all the models are applied to the scoring set and the labels are
combined by averaging the output (for regression) or voting (for classification).
3 Machine learning systems for Li-ion battery impedance
estimation
This study applies two machine learning systems (MLP neural network - Section
2.2 and gradient boosting - Section 2.3) in order to estimate the battery internal
impedance and compares the obtained results.
To perform the comparison between the two machine learning systems, a battery
testing database provided by NASA Ames Research Center [14] was used as data
set. The database comprises sensor monitoring data of Li-ion batteries mounted in
batches of 4 and running through 3 different operational profiles (charge, discharge
and impedance) at ambient temperatures of 4, 24 and 44 Celsius degrees (see Section
4).
From the analyzed data set, features were selected to characterize the battery
condition during a certain period of time (see Section 5). After the development
6

of the data set, machine learning systems are trained (see Section 6), Root Mean
Square Error (RMSE) performance metric is used to evaluate the obtained machine
learning systems (see Section 9) and, if the performance criteria are achieved, they
can be considered feasible solutions to estimate the battery internal impedance in
an online basis.
4 Data preparation for Li-ion battery impedance estimation
Battery impedance, which decreases over the working time of a battery, is an im-
portant and direct indicator for estimating battery state of charge (SOC). In online
or in-orbit applications, such as electric vehicles and satellites, the battery inter-
nal impedance measurement or monitoring is difficult [11]. It can be used charge
transfer resistance and electrolyte resistance extracted from EIS to estimate battery
capacity [14]. However, these features can only be obtained via offline tests under
the optimal measuring conditions and by using specialized and expensive equip-
ment for EIS measurements [5]. The results of the aging experiment showed that
increase in battery capacity loss or resistance in a lifetime is related to operating
conditions, such as voltage, current, and temperature. However, in practical appli-
cations, several characteristics, such as current and voltage, are controlled to meet
the load requirements of an associated circuit and thus cannot represent battery
aging [13].
4.1 Li-ion battery testing set information
The following sections apply machine learning techniques in Li-ion battery impedance
estimation using battery data provided by National Aeronautics and Space Admin-
istration (NASA) Ames Prognostics Center of Excellence [14], where 134 recharge-
able lithium-ion batteries were tested.
The laboratory setup and data recording were conducted by National Aero-
nautics and Space Administration (NASA) Ames Prognostics Center of Excellence
[14]. According to NASA experiment, the 134 rechargeable lithium-ion batteries
are organized in 34 battery sets. Each battery set contains the test data organized
according to Figure 2.
4.2 Laboratory setup
The experimental setup primarily consists of a set of Li-ion cells (which may reside
either inside or outside an environmental chamber), chargers, loads, EIS equipment
for battery health monitoring (BHM), a suite of sensors (voltage, current and tem-
perature), some custom switching circuitry, data acquisition system and a computer
for control and analysis. Figure 1 details the assembly of the equipment.
7

Fig. 1. Laboratory setup [14]
The cells are cycled through charge and discharge cycles under different load
and environmental conditions set by the electronic load and environmental chamber
respectively. Periodically EIS measurements will be taken to monitor the internal
condition of the battery using the BHM module. The DAQ system collects the
externally observable parameters from the sensors. The switching circuitry enables
the cells to be in the charge, discharge or EIS health monitoring state as dictated
by the algorithms running on the control computer [14].
4.3 Li-ion battery testing set information
The Li-ion batteries are organized in batches of 4 are run through 3 different op-
erational profiles (charge, discharge and impedance) at ambient temperatures of 4,
24 and 44 oC [14]:
1. Charge step: charging was carried out in a constant current (CC) mode at 1.5A
until the battery voltage reached 4.2V and then continued in a constant voltage
(CV) mode until the charge current dropped to 20mA
2. Discharge step: discharging was conducted in CC mode until the discharge volt-
age reached a predefined cutoff voltage. Fixed and variable load currents at 1,
2, and 4 Amps were used and the discharge runs were stopped at 2V, 2.2V,
2.5V or 2.7V
3. Impedance measurement: measurement was performed through an electrochem-
ical impedance spectroscopy (EIS) frequency sweep from 0.1 Hz to 5 kHz
Figure 2 details the battery data structure of the operational profiles.
4.4 Li-ion battery impedance measurement rectifier
In order to eliminate the noise generated by time-varying current passing through
an electro-chemical cell or battery due to load fluctuation, a filtering approach or an
8

Fig. 2. Data structure [14]
electronic cancellation technique shall be applied. In the laboratory setup described
above, it was used an electronic device.
Time-varying current flowing in a circuit which includes the cell/battery is
sensed externally to the cell/battery with a magnetically-coupled ac current probe
thereby producing an induced time-varying signal. This induced signal is amplified
to the level of the original time-varying current and applied to the cell/battery’s
terminals in phase-opposition to the original current. As a result, the component of
time-varying current flowing in the cell/battery’s external leads assumes an alter-
nate path around the cell/battery and is effectively canceled within the cell/battery
itself [2].
5 Features selection for Li-ion battery impedance estimation
In this study, the 134 rechargeable lithium-ion batteries are organized in 34 battery
sets. Each battery set contains the test data according to Figure 2. For each C-D
(charge and discharge) cycle, the following features were extracted [20]:
– F1: during charge cycle, time interval between the nominal voltage and the
cutoff voltage
– F2: during charge cycle, time interval between the nominal current and the
cutoff current
– F3: during discharge cycle, time interval between two predefined voltages
– F4: average temperature during the time interval F1
– F5: average temperature during the time interval F2
– F6: during discharge cycle, cutoff voltage
The historical set applied in the machine learning systems modeling includes
the six features (F1, F2, F3, F4, F5 and F6) and the label attribute which corre-
sponds to the rectified battery impedance. For each machine learning system, two
9

regression models are built in order to estimate the real and imaginary components
of the rectified impedance. Tables 1 and 2 are extractions of the historical sets
corresponding to the real and imaginary components which contain 20 records (C-D
cycles) for each one of the 34 battery sets (684 records - C-D cycles, in total).
Table 1. Historical set - impedance real component
cycle F1 F2 F3 F4 F5 F6 imp re
2008 4 1 9966.407 6422.609 472.313 27.0074 26.3008 2.4556 0.06175
2008 4 2 10226.375 6627.891 472.125 25.6742 26.5323 2.6321 0.05989
2008 4 3 10635.968 6528.063 472.344 25.6754 26.3325 2.5010 0.05919
Table 2. Historical set - impedance imaginary component
cycle F1 F2 F3 F4 F5 F6 imp img
2008 4 1 9966.407 6422.609 472.313 27.0074 26.3008 2.4556 -0.00096
2008 4 2 10226.375 6627.891 472.125 25.6742 26.5323 2.6321 -0.00112
2008 4 3 10635.968 6528.063 472.344 25.6754 26.3325 2.5010 -0.00105
6 Li-ion battery impedance data modeling
After the data preparation, two machine learning techniques (MLP neural network
- Section 2.2 and gradient boosting - Section 2.3) are applied in order to estimate
the battery internal impedance. For each applied technique, a few training cycles
are executed with the variation of the technique hyperparameters. Therefore, each
training cycle generates a regression model which can be compared with the other
obtained regression models by the application of the RMSE (Root Mean Square
Error) validation technique, according to Section 2.3.
7 MLP neural network with monotonicity constraints modeling
Tables 1 and 2 are used to build two regression models based on MLP neural
network technique (Section 2.2).
Multi Layer Perceptron neural network with monotonicity constraints imple-
ments one hidden-layer Multi Layer Perceptron neural network (MLP) models that
enforces monotonic relations on designated input variables. Each training cycle ap-
plies 10 or 20 ensemble members to fit and 1, 2, 3 or 4 hidden nodes in the hidden
layer.
The ensemble members to fit in each training cycle are obtained according to
Section 2.4. Each ensemble member contains a random subset of 70 percent of the
10

training set and, after generating the 10 or 20 regression models (depending on
the number of ensemble members to fit), the result corresponds to the mean value
obtained through the application of all regression models.
These training cycles with different combinations result into different MLP neu-
ral networks with monotonicity constraints. These different combinations of the
hyperparameters are described in Table 3.
Table 3. MLP neural network with monotonicity constraints hyperparameters
Hyperparameter Description Possible values
hidden1 number of hidden nodes in the first hidden layer 1, 2, 3, 4
n.ensemble number of ensemble members to fit 10, 20
monotone
column indices of covariates for which the
monotonicity constraint should hold
1
bag
logical variable indicating whether or not
bootstrap aggregation (bagging) should
be used
TRUE
iter.max
maximum number of iterations of
the optimization algorithm
500
7.1 MLP neural network modeling with monotonicity constraints and
10 ensemble members to fit
Estimative of real component of battery set impedance Using the back-
propagation algorithm to update the MLP neural network weights, monotonicity
constraints, 10 ensemble members to fit and using different numbers of hidden nodes
in the hidden layer (1, 2, 3 or 4), the real component of the battery impedance
through the cycles is according the Figure 3.
Estimative of imaginary component of battery set impedance Using the
backpropagation algorithm to update the MLP neural network weights, monotonic-
ity constraints, 10 ensemble members to fit and using different numbers of hidden
nodes in the hidden layer (1, 2, 3 or 4), the imaginary component of the battery
impedance through the cycles is according the Figure 4.
7.2 MLP neural network modeling with monotonicity constraints and
20 ensemble members to fit
Estimative of real component of battery set impedance Using the back-
propagation algorithm to update the MLP neural network weights, monotonicity
constraints, 20 ensemble members to fit and using different numbers of hidden nodes
in the hidden layer (1, 2, 3 or 4), the real component of the battery impedance
through the cycles is according the Figure 5.
11

Fig. 3. Estimated real component of battery set impedance using MLP neural network modeling
with monotonicity constraints and 10 ensemble members to fit
Fig. 4. Estimated imaginary component of battery set impedance using MLP neural network
modeling with monotonicity constraints and 10 ensemble members to fit
Estimative of imaginary component of battery set impedance Using the
backpropagation algorithm to update the MLP neural network weights, monotonic-
ity constraints, 20 ensemble members to fit and using different numbers of hidden
nodes in the hidden layer (1, 2, 3 or 4), the imaginary component of the battery
impedance through the cycles is according the Figure 6.
7.3 MLP neural network with monotonicity constraints model
validation
Applying the RMSE (Root Mean Square Error) validation technique, according to
Section 2.3, each model developed with the application of Multi Layer Perceptron
12

Fig. 5. Estimated real component of battery set impedance using MLP neural network modeling
with monotonicity constraints and 20 ensemble members to fit
Fig. 6. Estimated imaginary component of battery set impedance using MLP neural network
modeling with monotonicity constraints and 20 ensemble members to fit
technique in the estimative of real and imaginary components of the battery set
impedance was evaluated and the corresponding RMSE values are presented in
Table 4.
Increasing the number of ensemble members to fit, there was no impact on the
root square mean error for estimating the real component of battery set impedance.
However, a higher number of ensemble members to fit minimized the root square
mean error for estimating the imaginary component of battery set impedance.
According to Table 4, a higher number of hidden nodes on the hidden layer
minimizes the root square mean error for estimating the real and imaginary com-
ponents of battery set impedance.
13

Table 4. MLP neural network with monotonicity constraints model validation
Battery impedance Ensemble members to fit Hidden nodes on layer RMSE
real 20 1 0.0161220600
real 20 2 0.0124101700
real 20 3 0.0100222800
real 20 4 0.0094165090
imaginary 20 1 0.0011236520
imaginary 20 2 0.0007225610
imaginary 20 3 0.0004410664
imaginary 20 4 0.0003257367
The MLP models with the lower root square mean error in estimating the real
and imaginary components of battery set impedance have the configuration of 20
ensemble members to fit and 4 hidden nodes on the hidden layer and an RMSE
error of 0.0094165090 and 0.0003257367, respectively.
8 Gradient boosting modeling
Tables 1 and 2 are used to build two regression models based on gradient boosting
technique (Section 2.3). The training set is divided into training and validation sets.
The validation set is used to evolve the regression models during each gradient
boosting iteration. Each training cycle applies an ’eta’ equal to 0.20 or 0.60 to
control the learning rate and a subsample ratio of the training instances of 0.80,
0.85, 0.90 or 0.95.
These training cycles with different combinations result into different gradient
boosting models. These different combinations of the hyperparameters are described
in Table 5.
Table 5. Gradient boosting hyperparameters
Hyperparameter Description Possible values
objective objective function reg:linear
max depth maximum depth of a tree 10
eta
control the learning rate: scale
contribution of each tree by a factor
of 0 <eta <1
0.20, 0.60
col sample
subsample ratio of columns
when constructing each tree
0.80
ss sample
subsample ratio of the training instance.
0.5 means that xgboost randomly
collected half of the data to grow trees
and this will prevent overfitting
0.80, 0.85, 0.90, 0.95
eval metric evaluation metric per validation cycle root mean square error
nrounds the max number of validation cycles 200
14

The hyperparameter ’eta’ has an important role in the gradient boosting mod-
eling since it controls the learning rate. This parameter scales the contribution of
each tree by a factor between 0 and 1 when it is added to the current approximation.
It is used to prevent overfitting by making the boosting process more conservative.
Lower value for ’eta’ implies larger value for ’nrounds’: low ’eta’ value means model
more robust to overfitting but slower to compute.
8.1 Gradient boosting modeling with eta = 0.20
Estimating real component of battery set impedance Building an ensemble
of decision trees in which each decision tree is built in order to minimize the error
of the previous one (gradient boosting method), setting the maximum depth of the
trees to 10, applying the root mean square error method as a metric to evolve the
model with the validation set, controlling the learning rate through ’eta’ of 0.20
and subsetting the training set instances with distinct ratios (0.80, 0.85, 0.90 or
0.95), the real component of the battery impedance through the cycles is according
the Figure 7.
Fig. 7. Estimated real component of battery set impedance using gradient boosting modeling with
eta = 0.20
Estimating imaginary component of battery set impedance Building an
ensemble of decision trees in which each decision tree is built in order to minimize
the error of the previous one (gradient boosting method), setting the maximum
depth of the trees to 10, applying the root mean square error method as a metric
to evolve the model with the validation set, controlling the learning rate through
’eta’ of 0.20 and subsetting the training set instances with distinct ratios (0.80,
15

0.85, 0.90 or 0.95), the imaginary component of the battery impedance through the
cycles is according the Figure 8.
Fig. 8. Estimated imaginary component of battery set impedance using gradient boosting modeling
with eta = 0.20
8.2 Gradient boosting modeling with eta = 0.60
Estimating real component of battery set impedance Building an ensemble
of decision trees in which each decision tree is built in order to minimize the error
of the previous one (gradient boosting method), setting the maximum depth of the
trees to 10, applying the root mean square error method as a metric to evolve the
model with the validation set, controlling the learning rate through ’eta’ of 0.60
and subsetting the training set instances with distinct ratios (0.80, 0.85, 0.90 or
0.95), the real component of the battery impedance through the cycles is according
the Figure 9.
Estimating imaginary component of battery set impedance Building an
ensemble of decision trees in which each decision tree is built in order to minimize
the error of the previous one (gradient boosting method), setting the maximum
depth of the trees to 10, applying the root mean square error method as a metric
to evolve the model with the validation set, controlling the learning rate through
’eta’ of 0.60 and subsetting the training set instances with distinct ratios (0.80,
0.85, 0.90 or 0.95), the imaginary component of the battery impedance through the
cycles is according the Figure 10.
16

Fig. 9. Estimated real component of battery set impedance using gradient boosting modeling with
eta = 0.60
Fig. 10. Estimated imaginary component of battery set impedance using gradient boosting mod-
eling with eta = 0.60
8.3 Gradient boosting model validation
Applying the RMSE (Root Mean Square Error) validation technique, according to
Section 2.3, each model developed with the application of Gradient boosting tech-
nique in the estimating real and imaginary components of the battery set impedance
was evaluated and the corresponding RMSE values are presented in Table 6.
A lower ’eta’ minimizes the root square mean error for estimating the real and
imaginary components of battery set impedance. According to Table 6, increasing
the number of instances to be used in the training process, there was a decrease
on the root square mean error for estimating the real component of battery set
impedance. However, a higher number of instances to be used in the training process
17

Table 6. Gradient boosting model validation
Battery impedance ’eta’ - Learning rate Subsample ratio RMSE
real 0.20 0.80 0.0057317940
real 0.20 0.85 0.0044927730
real 0.20 0.90 0.0034616720
real 0.20 0.95 0.0028836380
imaginary 0.20 0.80 0.0003310718
imaginary 0.20 0.85 0.0006632989
imaginary 0.20 0.90 0.0005046742
imaginary 0.20 0.95 0.0001962272
maximized the root square mean error for estimating the imaginary component of
battery set impedance.
The Gradient boosting models with the lower root square mean error in es-
timating the real and imaginary components of battery set impedance have the
configuration of ’eta’=0.20 and a subsample ratio of 0.95 and an RMSE error of
0.0028836380 and 0.0001962272, respectively.
9 Li-ion battery impedance model validation
According to RMSE validation technique (see Section 2.3), the Multi Layer Per-
ceptron model with the higher performance (20 ensemble members to fit and 4
hidden nodes on the hidden layer) and the Gradient Tree Boosting model with the
higher performance (’eta’=0.20 and a subsample ratio of 0.95) are compared in 11
regarding the estimation of the real and imaginary components of the battery set
impedance.
Fig. 11. RMSE State of charge (SOC) model validation
18

10 Conclusion
In this study, two distinct machine learning approaches were applied in order to
estimate the battery impedance of satellite Li-ion battery sets which is used in the
determination of their state of charge (SOC) (see Section 2.4). The obtained RMSE
(Root Mean Square Error) of both approaches has demonstrated the feasibility of
such machine learning systems in estimating Li-ion battery sets battery impedance
in cases where an error below 0.10 is acceptable.
However, once the obtained RMSE (Root Mean Square Error) of the Gradient
Tree Boosting model is lower in estimating both real and imaginary components
of the battery impedance, this approach is recommended over the Multi Layer
Perceptron model.
Acknowledgements
M.G.Q thanks FAPESP (Grant 2011/18496-7) and CNPq (Grant 310908/2015-9).
References
1. J. A. Aslam, R. A. Popa, and R. L. Rivest. On estimating the size and confidence of a
statistical audit. Proceedings of the Electronic Voting Technology Workshop, 7, 2007.
2. K. S. Champlin. Method and apparatus for suppressing time-varying signals in batteries
undergoing charging or discharging, 11 1992.
3. Y. Chang, W. The state of charge estimating methods for battery: A review. International
Scholarly Research Notices Applied Mathematics, 2013:1–7, 2013.
4. M. Coleman, C. K. Lee, C. Zhu, and W. G. Hurley. State of-charge determination from emf
voltage estimation: using impedance, terminal voltage, and current for lead-acid and lithium-
ion batteries. IEEE Transactions on Industrial Electronics, 54:25502557, 2007.
5. M. Dalal, J. Ma, and D. He. Lithium-ion battery life prognostic health management system
using particle filtering framework. Proc. Inst. Mech. Eng. Part O J. Risk, 225:8190, 2011.
6. G. J. Dudley. Lithium-ion batteries for space. Proceedings of the Fifth European Space Power
Conference, page 17, 1998.
7. T. Hastie, R. Tibshirani, and J. H. Friedman. The elements of statistical learning: data mining,
inference, and prediction, volume 02. Springer, USA, 2009.
8. S. Haykin. Neural networks - A comprehensive foundation, volume 02. Cambridge University
Press, USA, 1999.
9. R. J. Hyndman and A. B. Koehler. Another look at measures of forecast accuracy. Interna-
tional Journal of Forecasting, 04:679688, 2006.
10. R. Li, J. F. Wu, H. Y. Wang, and G. C. Li. Prediction of state of charge of lithium-ion
rechargeable battery with electrochemical impedance spectroscopy theory. Proceedings of the
5th IEEE Conference on Industrial Electronics and Applications, page 684688, 2010.
11. D. Liu, H. Wang, Y. Peng, W. Xie, and H. Liao. Satellite lithium-ion battery remaining cycle
life prediction with novel indirect health indicator extraction. Energies, 6:36543668, 2013.
12. J. Liu and G. Wang. A multi-step predictor with a variable input pattern for system state
forecasting. Mechanical Systems and Signal Processing, 23:15861599, 2009.
13. M. Parviz and S. Moin. Boosting approach for score level fusion in multimodal biometrics
based on auc maximization. J. Inf. Hiding Multimed. Signal Process, 2:5159, 2011.
19

14. B. Saha and K. Goebel. Battery data set, 2007. NASA Ames Research Center, Moffett Field,
CA.
15. S. Sato and A. Kawamura. A new estimation method of state of charge using terminal voltage
and internal resistance for lead acid battery. Proceedings of the Power Conversion Conference,
page 565570, 2002.
16. M. Schwabacher. A survey of data-driven prognostics. AIAA Meeting Papers, 2005.
17. D. Wang, G. Li, and Y. Pan. The technology of lithium-ion batteries for spacecraft application.
Aerospace Shanghai, 4:5459, 2000.
18. N. Watrin, B. Blunier, and A. Miraoui. Review of adaptive systems for lithium batteries state-
of-charge and state-of-health estimation. Proceedings of IEEE Transportation Electrification
Conference and Expo, pages 1–6, 2012.
19. H. Zhang and Z. Zhang. Feedforward networks with monotone constraints. International Joint
Conference on Neural Networks, 03:1820–1823, 1999.
20. J. Zhang and J. Lee. A review on prognostics and health monitoring of li-ion battery. Journal
of Power Sources, 196:60076014, 2011.
Authors
Thiago Donato graduated at Electrical Engineering from Fed-
eral University of Itajub - UNIFEI (2006) and has worked dur-
ing five years in private companies ( EMBRAER - aircraft man-
ufacturer company - and TOTVS - software company) develop-
ing solutions which apply machine learning techniques in the
resolution of major issues. Donato is enrolled in master’s at
Computer Science from National Space Research Institute -
INPE (2018). Has experience in Machine Learning, acting on
the following subjects: data and text preparation and mining,
neural network and other machine learning classification tech-
niques.
Marcos G. Quiles is an Associate Professor at the Depart-
ment of Science and Technology, Federal University of So Paulo,
Brazil. He received the BS degree, with honors, in 2003 from the
State University of Londrina, Brazil, and a Ph.D. degree from
the University of So Paulo, Brazil, in 2009, both in Computer
Science. From January to July of 2008, Quiles was a Visiting
Scholar in the Perception and Neurodynamics Lab at The Ohio
State University. From January to December of 2017, Quiles
was an Academic Visitor at the University of York, York-UK.
He was awarded a Brazilian research productivity fellowship
from the Brazilian National Research Council (CNPq). His re-
search interests include nature-inspired computing, machine learning, complex net-
works, and their applications in interdisciplinary problems.
20

Machine learning systems based on xgBoost and MLP neural network applied in satellite lithium-ion battery sets impedance estimation

More Related Content

Similar to Machine learning systems based on xgBoost and MLP neural network applied in satellite lithium-ion battery sets impedance estimation (20)

Recently uploaded (20)

Machine learning systems based on xgBoost and MLP neural network applied in satellite lithium-ion battery sets impedance estimation