Measurement 183 (2021) 109838

Contents lists available at ScienceDirect

Measurement
journal homepage: www.elsevier.com/locate/measurement

Modeling of a DC-DC bidirectional converter used in mild hybrid electric

vehicles from measurements
Gabriel Rojas-Dueñas , Jordi Roger Riba *, Manuel Moreno-Eguilaz
Universitat Politècnica de Catalunya, Electrical Engineering Department, Rambla Sant Nebridi 22, 08222 Terrassa, Catalunya, Spain

A R T I C L E I N F O A B S T R A C T

Keywords: This paper presents a non-intrusive approach for modeling a bidirectional DC-DC converter used in mild hybrid
DC-DC bidirectional converter electric vehicles. A black-box identification methodology is proposed to find a model based on the data acquired
Mild hybrid electric vehicle from the input/output terminals. Measured data include the steady state and transient response, and different
Deep learning
operating conditions of the DC-DC converter, including the buck and boost modes. A deep learning architecture
Modeling
Neural network
based on a long-short-term memory neural network (LSTM-NN) is applied. The trained network is tested under a
set of operating points different from those used during the training stage. The proposed method is compared
with three black-box modeling techniques commonly used in power converters, proving its superior perfor­
mance. Results presented in this paper indicate that the proposed model is able to replicate the behavior of the
bidirectional converter without a priori knowledge of the converter circuitry. This approach can also be applied
to other power devices.

1. Introduction automobile manufacturers acquire these electrical components from

external suppliers. Manufacturers of electrical components often do not
Over the last years, the automotive industry has been moving to­ provide enough information on the dynamics, operating points or to­
wards more efficient vehicles in order to achieve stringent CO2 emission pology of their products [8]. This presents a challenge for MHEV engi­
targets [1]. To achieve this goal automobiles must be electrified, and in neers because they require accurate models of the electrical equipment
particular the powertrains [2,3]. Although full electric vehicles are for designing, optimization, control, fault diagnosis [9] and useful life
possibly the best solution because they allow the highest reduction in estimation purposes. Thus, it is important to model and identify these
emission of pollutants [2], a wide expansion is not yet possible since elements to fulfill the aforementioned requirements. Black-box models
they are expensive compared to conventional ones and the lack of can be useful in addressing this problem, as they do not require prior
required infrastructures. Therefore, hybrid electric vehicles arise as a knowledge of the topology of the device to be identified or modeled.
solution for the transition from the fuel based technology to the full Different authors have proposed modeling approaches of the 48 V and
electric vehicles. 12 V batteries used in HEVs [6,10]. However, just some few studies have
Hybrid electric vehicles (HEVs) consist of two or more power sour­ analyzed the identification of bidirectional DC-DC converters. There­
ces, in which one of them is characterized by being electrical and fore, this paper attempts to fill this gap, as it seeks to obtain an accurate
reversible [3]. This paper analyzes a mild hybrid electric vehicle black-box model of the 48 V –to- 12 V bidirectional power converter
(MHEV), this technology generating around 15 % less emissions than used in MHEVs based on an offline black-box identification methodol­
conventional vehicles [4]. MHEVs reduce the fuel consumption by using ogy that uses a deep learning neural network architecture.
regenerative braking and by moving the operating points of the thermal Depending on the information available, different approaches can be
engine to more efficient areas [5] with the help of the electrical motor. applied to model power converters. A white-box parameter identifica­
The deceleration energy is stored in an electrical storage system [6], tion approach is used when the topology of the DC-DC converter is
which uses a 48 V –to- 12 V dual-voltage architecture consisting of two known, although the parameter values are unknown. These problems
batteries and a bidirectional DC-DC power converter [7]. In most cases, are usually solved by solving the analytical equations describing the

* Corresponding author.at: Universitat Politècnica de Catalunya, Electrical Engineering Department, Rambla Sant Nebridi 22, 08222 Terrassa, Catalunya, Spain.
E-mail addresses: gabriel.esteban.rojas@upc.edu (G. Rojas-Dueñas), jordi.riba-ruiz@upc.edu (J. Roger Riba), manuel.moreno.eguilaz@upc.edu (M. Moreno-
Eguilaz).

https://doi.org/10.1016/j.measurement.2021.109838
Received 3 May 2021; Received in revised form 22 June 2021; Accepted 26 June 2021
Available online 10 July 2021
0263-2241/© 2021 Elsevier Ltd. All rights reserved.
G. Rojas-Dueñas et al. Measurement 183 (2021) 109838

dynamics of the circuit and using optimization methods [11]. Alterna­

tively, black-box modeling methods, as in this paper, have a higher
complexity due to the lack of information about the physical charac­
teristics of the power converter. In this case, the only available infor­
mation are the currents and voltages at the input and output terminals
[12]. These modeling approaches prevent manufacturers of disclosing
the information of their products [13].
In [14] different black-box modeling techniques of power converters
are reviewed, including the Hammerstein and Steiglitz IIR models,
which show a good performance. However, these are online estimation
approaches, requiring an external change of the duty cycle, leading to a
higher complexity and limited robustness because these methods require
a specific model for each operating point [15]. In [16] the frequency
response of the power converter is obtained, which is used to generate a
transfer function able to replicate its behavior. Despite of allowing to
replicate the dynamics of the converter, this method might be inaccurate
when the converter works in the discontinuous conduction mode. In
[12] a polytopic model is proposed, which obtains the G-parameters of
the two-port network (converter) and calculates the linear response at a
certain operating point. This method presents accurate results, but it is
limited to the analyzed conduction mode of the DC-DC converter, thus
failing to estimate changes between the continuous and discontinuous
conduction modes. In [17] a black-box model of a buck converter is
obtained by using nonlinear autoregressive exogenous neural networks
(NARX-NNs). An offline identification is performed and the results show
a good resemblance between the estimation and the measured output of
the converter. Nevertheless, the computational burden is high and the
inputs of the black-box are not the same as the ones of the converter. In
[18] an identification method based on recursive neural networks is
presented, however, it fails to replicate with high-fidelity the behavior of
the converter and it just uses data gathered from simulation. The authors Fig. 1. Proposed measurement-based methodology.
of [19] propose a dynamic system identification method of a DC-DC
bidirectional converter used in automobiles. It is based on three with in this work, including the suggested design and the tuning of the
different techniques, each one being applied in a certain condition of the hyper-parameters of the neural network. In Section 4, the experimental
converter. It uses the state space averaging method, which is also setup and measurement strategy are presented and in Section 5 the
applied in [20] to obtain the model of a buck converter. Although the attained results are shown and analyzed. Finally, Section 6 concludes the
error obtained is very low, it was just tested under buck mode, and the paper.
operating point of the converter must be known in order to apply one of
the three techniques. In [9] the authors use neural networks for pre­ 2. Methodology
dictive control and fault diagnosis of power converters, respectively.
However, a model that replicates the behavior of the converter is not This paper uses a deep learning approach to model a 48 V –to- 12 V
obtained. DC-DC bidirectional converter used in mild hybrid electric vehicles. The
The methodology proposed in this study is based on offline and non- proposed methodology considers three stages, which comprise the
intrusive measurements of the DC-DC bidirectional converter. Some of acquisition of the experimental data from the converter, the design and
them are used to train a long short-term memory neural network (LSTM- implementation of the LSTM-NN and the validation of the model.
NN), while the others test the accuracy of it. An experimental setup Fig. 1 shows the general procedure to acquire the experimental data,
capable of obtaining the needed datasets is implemented. It automati­ train and validate the model, where the three stages are clearly defined.
cally acquires different conditions of the power converter in order to The first stage of the proposed methodology comprises the data
provide enough information to the neural network to generate an acquisition of the bidirectional DC-DC converter. To this end, it is
adequate model. One of the main advantages of using this deep learning necessary to define the equipment to be used and the data to be
approach is that the neural network is able to learn the long term de­ measured in the laboratory. An automatic data acquisition system was
pendencies of the converter behavior as well as to predict the future programmed for this purpose, which is able to store a predefined num­
outputs based on the past and actual states of the LSTM-NN. Another ber of experiments. The gathered data is resized and filtered to increase
advantage is the capacity of replicating the steady-state and transient the model accuracy and decrease the training time. Finally, the resulting
response characteristics of the converter using only one model, and not dataset is divided in the training, validation and test sets. A more
several, as in other works. Thus, different operating points of the con­ detailed explanation of the experimental stage is given in Section 5.
verter can be replicated by means of a single trained LSTM-NN. It is The LSTM-NN stage shown in Fig. 1 defines the neural network and
worth noting that the only needed measurements are non-intrusive and selects the training parameters. First, the inputs and outputs of the NN
they are made at the input and output terminals of the converter, which are specified, in this case corresponding to the measurements at the
prevents an alteration of the inner elements of the converter. Also, un­ input and output terminals (voltages and currents) of the bidirectional
like other studies [18], the experimental setup allows a robust modeling converter. Next, the number of hidden layers, the fixed parameters and
of the bidirectional DC-DC converter. The entire MHEV electrical system the hyper-parameters to be tuned are defined. In the last step of this
can be simulated if an existing battery model [21,22] is added to the stage, the hyper-parameters are tuned by means of an optimization al­
bidirectional converter. gorithm, which is further explained in Subsection 3.2. Next, the LSTM-
Section 2 details the methodology proposed in this paper as well as NN is trained using the optimized hyper-parameters and the training
the electrical system of MHEVs. Section 3 details the LSTM-NN dealt and validation datasets. To determine the accuracy of the model, the

2
G. Rojas-Dueñas et al. Measurement 183 (2021) 109838

Fig. 2. MHEV electrical system.

neural network is tested under new conditions, and the predicted out­
puts are compared to the actual measured outputs of the bidirectional
converter. Finally, the root mean squared error (RMSE) and the coeffi­
cient of determination (R2) for the two predicted signals are calculated.
Fig. 3. LSTM cell unit.

2.1. Mild hybrid electrical vehicle (MHEV) modeling time-series problems with good accuracy. However, their
main problem is the difficulty of learning long term dependencies of the
The system to be evaluated in this paper is a MHEV, this topology data because of the gradient decay. LSTM-NNs overcome this problem
being used by major manufacturers such as Audi, Toyota, Nissan or [26]. This type of network is able of discarding unrelated information
Honda, among others [23]. This type of vehicle improves fuel efficiency during the training process, leading to the preservation of the back­
by supplying power to the electric motor in certain situations or by propagation error over time. The weights of the neural network are
applying regenerative braking. MHEVs require a higher voltage (48 V) updated based on the memory units of the LSTM cells, which are not
to feed the electrical systems of the vehicle in order to reduce conduction affected by external perturbations [27]. A great advantage of LSTM-
losses [5,24]. The MHEV is a transition technology that is placed in the NNs, compared to traditional RNNs, is that they keep learning at
middle point between the conventional 12 V supply system and the 48 V every time step, even in case of numerous time steps.
system proposed by some manufacturers of premier cars [7]. The fundamental unit of the LSTM-NN is the cell, which includes
This paper focuses on the dual-voltage architecture that considers memory blocks that store, write, read and erase information during the
two batteries and a DC-DC bidirectional converter, the last one being the training process. This is accomplished thanks to the presence of gates,
device to be modeled. Understanding the working principle and the that allow the network to update the weights of the neurons at each
topology of the electrical system of the MHEV is fundamental in order to iteration [28].
obtain an accurate model of the power converter. Fig. 3 shows the architecture of LSTM cells, where c refers to the cell
Fig. 2 shows the dual architecture of this electrical system and its memory state, h is the hidden state, x is the input value and σ represents
main components. It can be seen that at the 48 V side there is a battery, a the sigmoid function. Subscript t refers to the time step of the cell, since
starter generator, a drive (DC-AC or AC-DC depending whether the there are as many cells as time steps in the time series. The LSTM is a
machine works as a motor or as a generator, respectively) and loads, block where different cells are connected sequentially, so that each cell
whereas at the 12 V side there is a battery and a set of low voltage loads. generates a future output depending on the past values. Fig. 3 shows the
Fig. 2 does not specify the topology of the DC-DC converter since it may input, forget, output and input modulation gates of a LSTM cell. The first
change depending on the manufacturer of the automobile and its re­ three are controlled by the sigmoid function σ , and they generate an
quirements [7]. However, for the purpose of this paper this is indif­ output between 0 and 1 based on the significance of the ht-1 input value
ferent, since the proposed black-box approach is able to model the [28], whereas the input modulation gate uses the ‘tanh’ activation
bidirectional converter despite of its topology. function. Pointwise operations are applied inside the cell in order to
Each battery is in charge of stabilizing its own network. When the connect the four gates and to obtain the output values. The interaction
belt-driven starter generator (BISG) is operating as a generator, the between the gates and the capacity of forgetting irrelevant information
converter works in the buck mode, the two batteries are charged and the allows the network learning long range temporal dependencies [9]. It is
power is supplied to the loads. When the BISG operates as a motor, the desirable for the problem studied in this paper, because the outputs of a
DC/DC converter operates in the boost mode and the power flows from DC-DC converter depend on the previous states [29].
the 12 V side to the 48 V side [5]. Some manufacturers include a starter This architecture allows calculating the output values just by
on the low voltage side in order to start the motor in case of failure of the applying mathematical calculations to the input data. Eq. (1) presents
48 V network. the operations that take place at the different gates of the LSTM cell.

3. Modeling of MHEV DC-DC converter i = S(W

( i xt + Ri ht− 1 + bi ))
f = S Wf xt + Rf ht− 1 + bf
(1)
o = S(W( o xt + Ro ht− 1 + bo ) )
The black-box modeling methodology presented in this paper is
g = tanh Wg xt + Rg ht− 1 + bg
based on a LSTM-NN. This is a type of recursive neural network
commonly applied in time-series related problems [25]. It uses the where W refers to the input weights, R to the recurrent weights, b to
actual and predicted values at the input and output terminals of the the bias, the subscript to the respective gate and the S() function rep­
power converter to predict the future output values. This section pre­ resents the sigmoid function applied to each gate. Then, the cell memory
sents a general overview of the well-known LSTM-NN architecture, next state and the hidden state are calculated as follows,
a specific LSTM-NN topology for the analyzed problem is presented, and
ct = f ⋅ct− 1 + i⋅g (2)
finally, a strategy to tune the hyper-parameters is developed.
ht = o⋅tanh(ct ) (3)
3.1. Long short term memory neural networks (LSTM-NN)
Finally, by using the updated hidden states, it is possible to calculate
the predicted output values as,
Traditional recurrent neural networks (RNN) are applied for

3
G. Rojas-Dueñas et al. Measurement 183 (2021) 109838

yt = Wfc ht + bfc (4) Table 1

Lm5170-Q1 Converter Specifications.
where Wfc and bfc are the weights and bias of the fully-connected
Parameter Value
layer, respectively.
Input voltage range 6 V – 75 V
Typical input voltaje 48 V
3.2. Training of the LSTM-NN Output voltage range 3 V – 48 V
Typical output voltage 12 V
Maximum power 720 W
To define the characteristics of the LSTM-NN and the tuning strategy, Switching frequency 100 kHz
it is necessary to establish the inputs and outputs of the black-box model. Peak efficiency > 97 %
Considering that the only variables that can be measured in real situa­
tions are the inputs and outputs of the DC-DC bidirectional converter,
the selected inputs of the black-box are the voltage and current at the 48 Table 2
V side, while the outputs are the voltage and current at the 12 V side. Battery Specifications.
However, the inputs and outputs might be in the opposite way if the 48 V
Parameter 48 V Battery 12 V Battery
is defined as the output.
After understanding the nature of MHEVs bidirectional converters, Nominal discharge current 6.28 A Not specified
Maximum discharge current 50 A 440 A
the working principle of the LSTM-NN and defining the inputs and Nominal charge current 6.28 A Not specified
outputs of the system, the next step is to choose the principal charac­ Maximum charge current 12 A Not specified
teristics of the LSTM-NN in order to model accurately the behavior of the Rated capacity 31.9 Ah 64 Ah
DC-DC converter. According to the LSTM-NN topologies presented in
[30], and given the complexity of the analyzed dataset, just one hidden
training of the LSTM-NN.
layer is required to obtain precise results after the training process.
To generate the aforementioned dataset, it is important to follow the
When adding extra layers to this type of NN, the training process de­
steps of the proposed methodology, as shown in Fig. 1. Therefore, the
mands more computational resources.
first step of the experimental stage is to choose an adequate set of bat­
The process of tuning the hyper-parameters is done by means of a
teries and a proper converter. These elements were selected considering
Bayesian optimization algorithm (BOA). It finds an optimal combination
the characteristics, ratings and specifications of the elements that can be
of hyper-parameters by training a small number of NNs. The algorithm
found in commercial MHEVs [23]. The chosen converter was the
trains the NN with an initial set of hyper-parameters, and based on the
LM5170-Q1 48 V -to- 12 V DC-DC buck-boost converter manufactured
points close to the minimum, the hyper-parameters are changed at each
by Texas Instruments. This converter is suitable for real MHEV appli­
iteration [31]. The objective function is defined as,
cations and it can be controlled externally by means of its current or a
x* = argmaxf (x) (5) voltage source. The evaluation module of the converter was used to
x∈X
perform the experiments, since it allows the user to modify different
where f(x) is a surrogate function generated by sampled points, x* parameters based on the required measurements. The bidirectional
refers to the optimal set of hyper-parameters and X to the space of converter has an inner current control loop and an outer voltage control
possible values. Given the surrogate function, the idea is to reduce the loop. The first one receives two external signals that determine the di­
uncertainty of the model by evaluating different LSTM-NNs with rection and magnitude of the current flowing through the converter. It
different sets of hyper-parameters. A good balance between exploration uses a Type-II compensator to stabilize the loop for the buck and boost
and exploitation of the candidate points is needed in the optimization operations. The outer voltage loop is controlled by a digital circuit and
process in order to have a high prediction accuracy with small uncer­ regulates the output voltage of the converter. Both loops are inter­
tainty [32]. connected, since the output error signal of the voltage loop is an input
Since the problem studied in this paper uses a large dataset, it is not signal for the current loop. Table 1 shows the characteristics of the
efficient to tune several hyper-parameters. Therefore, just four of them converter.
were chosen, based on how they affect the training process and model The MD-BB0 Li-Ion 48 V battery pack from Torrot was used, which is
accuracy. These hyper-parameters are the hidden layer neurons,
learning rate, gradient decay factor (GDF) and the L2 regularization
value. The first one determines the number of interconnections within
the NN and its complexity, the second one refers to the size of the step
that the training algorithm takes in each iteration. The last two hyper-
parameters are associated to the solver, which in this case is the
‘adam’. This optimizer is meant to solve a non-convex optimization
problem that updates the weights of the neurons at each iteration in
order to find a local minimum point [33]. The problem is unconstrained
and the convergence is affected by the selected hyper-parameters. The
GDF represents the decay rate and the L2 regularization parameter is the
weight decay of the training process. These variables help the network to
avoid overfitting.

4. Experimental setup and validation

The experimental setup pursues replicating the voltage conversion

stage that occurs in MHEVs. To this end, the setup shown in Fig. 2 was
implemented in the laboratory. It includes three main elements, the 48 V
battery, the DC-DC bidirectional power converter and the 12 V battery.
The objective is to obtain a rich dataset based on the measurements at
the input and output terminals of the converter, allowing an accurate Fig. 4. Experimental setup.

4
G. Rojas-Dueñas et al. Measurement 183 (2021) 109838

intended for automotive applications. Despite it is not specifically

designed for MHEVs, the characteristics shown in Table 2 are quite
similar to the 48 V batteries of commercially available MHEVs. On the
other hand, the T4 12 V lead acid battery from Augi was used, which is
widely applied in vehicles.
Table 2 summarizes the specifications of the batteries dealt with in
this work.
After selecting the equipment, the following step is to define an
experimental setup based on the type of data needed to train and test the
LSTM-NN. As already mentioned, the training stage of the neural
network requires a rich dataset that includes different operating points
and transient responses of the bidirectional converter. Thus, several
conditions must be measured based on the maximum rating and
response time of the converter. An automatic acquisition system was
Fig. 5. Experimental setup of the MHEV electrical system.
implemented in order to collect the training and test datasets in a fast
and efficient way.
Fig. 4 shows the schematic of the setup implemented in the labora­
tory where the oscilloscope and the bidirectional converter are centrally
controlled by a computer and a DAQ device. A Python code synchronizes
the control signals of the converter and the data acquisition system. This
code sends a current command to the control loops of the bidirectional
converter. Based on this value, it adjusts the measurement settings of the
oscilloscope (trigger, coupling and time scale). The maximum and
minimum values of the current command are fixed based on the power
rating of the batteries and the converter. The number of iterations is
defined by the user and indicates the number of experiments of the
dataset. Each experiment contains five signals: time, input voltage, input
current, output voltage and output current of the bidirectional con­
verter. It is not necessary to connect loads to any of the batteries, as the
converter provides the direction and magnitude of the current.
It is necessary to pre-process the input data in order to filter the noise
and to avoid overfitting, thus increasing the training speed of the neural
network. The pre-processing consists of an interpolation and the appli­
cation of a simple moving average filter. First, the interpolation is
applied to reduce the number of points of each experiment of the
dataset. Next, the moving average filter is applied to eliminate the high-
frequency noise from the measurements. The number of points taken in
the moving average is chosen in order to keep the main characteristics of
the transient response. The last step of this stage is to divide the pro­
cessed data set into the training, validation and test sets. This partition is
done based on the information available on the dataset, because it Fig. 6. Converter measurements per experiment. (a) Input voltage, (b) Input
directly affects the results of the training of the LSTM-NN. The same pre- current, (c) Output voltage and (d) Output current.
processing strategy must be applied to these three subsets.
technical literature, that is, NARX neural networks [17], polytopic
4.1. Validation modeling [34] and the state-space averaging method [19,35]. These are
methods that show accurate results in identifying the dynamics of DC-
The 48 V –to- 12 V DC-DC bidirectional converter model is validated DC converters and are comparable to the model proposed in this
by the test data set to predict the output values provided by the LSTM- paper, since they require offline measurements. It is noted that a basic
NN, which are compared against the actual values, and the error be­ topology of a buck-boost converter was modeled in case of the state-
tween both signals is calculated. space average model.
After pre-processing the raw acquired experimental data, the dataset Fig. 5 shows the experimental setup implemented in the laboratory,
was partitioned in three different subsets as shown in Fig. 1. The third from which the experimental data was acquired. This setup has the same
set, i.e., the test dataset, is used to validate the LSTM-NN by testing the configuration as that in Fig. 4.
obtained model under conditions other than those of training. This is The equipment used to acquire the data consists of a 4 channel
done to verify the robustness of the model and that the trained neural oscilloscope (Tektronix MDO3024 200 MHz 2.5 GS/s; Tektronix, Bea­
network does not have overfitting issues. The accuracy of the predictions verton, OR, USA), two high-frequency current probes (Tektronix
is calculated by means of the root mean squared error and the coefficient TCP0030A 0.001–20 A 120 MHz; Tektronix, Beaverton, OR, USA) and
of determination, which are two indicators used to evaluate modeling two high-frequency voltage probes (Tektronix THDP200; Tektronix,
and fitting problems [30]. RMSE values as low as possible and R2 values Beaverton, OR, USA). A total of 1000 different experiments were per­
close to 1 are desirable. formed by controlling the current flowing through the power converter.
The sampling frequency of the oscilloscope was set to 2.5 MHz with a
5. Experimental results length of 105 points, resulting in a period of 4 ms. The pre-processing
consisted of reducing the number of points of every experiment to
This section presents the modeling results of a real DC-DC buck-boost 5000 and then applying a moving average filter comprising 10 points.
converter used in MHEVs based on experimental data. The results ob­ Fig. 6 presents an example of the four signals acquired in each
tained are compared with three black-box methods found in the

5
G. Rojas-Dueñas et al. Measurement 183 (2021) 109838

Table 3
Bayesian Optimization Range.
Hyper-parameter Minimum Maximum

Neurons 5 150
Learning rate (LR) 0.001 0.1
Gradient decay factor (GDF) 0.8 0.99
L2 regularization value 1e-10 1e-2

Fig. 7. Training progress.

Table 4
Results.
Method RMSE R2 Time
elapsed
Iin Vout Iin Vout

LSTM-NN 0.1204 0.0023 0.9995 0.9998 274 s

NARX-NN [17] 1.4043 0.0067 0.9761 0.9877 12458 s
Polytopic model [34] 1.3740 0.0126 0.9522 0.9565 26 s
State-space averaging 1.4410 0.0222 0.9465 0.9539 312 s
method [19,35]

Fig. 9. Scatter plot comparing measured and estimated signals. (a) output
voltage (LSTM-NN), (b) output current (LSTM-NN), (c) output voltage (NARX),
(d) output current (NARX), (e) output voltage (Polytopic), (f) output current
Fig. 8. Transient response measurement and estimation of (a) Output voltage (Polytopic), (g) output voltage (State space), (h) output current (State space).
(boost mode to buck mode), (b) Output current (boost mode to buck mode), (c)
Output voltage (buck mode to boost mode), (d) Output current (buck mode to
The following step consists of validating the NN when tested under
boost mode).
different conditions than the ones used during the training stage.
Table 4 shows the RMSE and R2 values obtained after evaluating 125
experiment. In this case it is well appreciated the transient response of
experiments from the test dataset. It also compares these values to the
the bidirectional converter.
ones obtained by means of the predictions made using the other three
A total of 750 experiments was assigned to the training dataset,
methods. The proposed methodology outperforms the other three
whereas the validation and test datasets contain 125 experiments each.
techniques because the RMSE is lower and the R2 coefficient is closer to
The next step is to tune the hyper-parameters of the NN.
1.
Table 3 presents the minimum and maximum bounds of the variables
To better understand the accuracy of the proposed model, Fig. shows
to be tuned by the Bayesian optimization algorithm. A total number of
the actual and predicted outputs of two experiments of the test dataset.
30 NNs were trained by the BOA, each training process run for 100
Fig. a) and b) show an experiment in which the converter changes its
epochs.
operating mode from boost to buck, whereas in Fig. c) and d) the
The BOA took 12 min to train the 30 neural networks. It was found
operating mode changes in the opposite way. In both cases, the esti­
that the optimal point has the following hyper-parameters: 29 neurons,
mation of the LSTM-NN is close to the measurements. Fig. 8 also show
LR = 0.03426, GDF = 0.974 and L2 = 0.00169. Subsequently, the LSTM-
that the accuracy of the transient response is better than those estimated
NN was trained with these hyper-parameters for 1000 epochs using the
by the other three models. The LSTM-NN model accurately replicates the
‘adam’ solver.
steady state and transient response values, while the other methods
Fig. 7 shows the evolution of the training progress.
exhibit some discrepancies. It is worthy to note that the estimation of the

6
G. Rojas-Dueñas et al. Measurement 183 (2021) 109838

current at the 12 V side of the power converter is slightly less accurate [4] M. Awadallah, P. Tawadros, P. Walker, N. Zhang, Comparative fuel economy, cost
and emissions analysis of a novel mild hybrid and conventional vehicles, Proc. Inst.
than the estimation of the voltage. This might be because the transient
Mech. Eng. Part D J. Automob. Eng., Nov. 232 (13) (2018) 1846–1862.
response of the output current is wider than the one for the voltage, [5] S. Nazari, J. Siegel, A. Stefanopoulou, Optimal Energy Management for a Mild
depending on the experiment. Hybrid Vehicle with Electric and Hybrid Engine Boosting Systems, IEEE Trans.
Fig. 9 plot the estimation done by the four methods against the Veh. Technol. 68 (4) (2019) 3386–3399.
[6] A. Lievre, A. Sari, P. Venet, A. Hijazi, M. Ouattara-Brigaudet, S. Pelissier, Practical
experimental measurements for different operating. Such results prove online estimation of lithium-ion battery apparent series resistance for mild hybrid
the robustness of the proposed model since the slope is very close to vehicles, IEEE Trans. Veh. Technol. 65 (6) (2016) 4505–4511.
unity. The proposed method outperforms the other three approaches, [7] T. Kim, S. Kwak, A Flexible Voltage Bus Converter for the 48-/12-V Dual Supply
System in Electrified Vehicles, IEEE Trans. Veh. Technol. 66 (3) (2017) 2010–2018.
since the dispersion of the points is much smaller and the coefficients of [8] V. Valdivia, Behavioral Modeling and Identification of Power Electronics
determination are greater. Converters and Subsystems Based on Transient Response, Universidad Carlos III de
The results presented in this section proved the effectiveness of the Madrid (2013).
[9] Z.Y. Xue, K.S. Xiahou, M.S. Li, T.Y. Ji, Q.H. Wu, “Diagnosis of Multiple Open-
modeling approach proposed in the paper, as well as the accurateness in Circuit Switch Faults Based on Long Short-Term Memory Network for DFIG-Based
comparison with other methods found in the technical literature. Wind Turbine Systems, IEEE J. Emerg. Sel. Top. Power Electron. 8 (3) (2020)
2600–2610.
[10] R. Xiong, J. Cao, Q. Yu, H. He, F. Sun, Critical Review on the Battery State of
6. Conclusion Charge Estimation Methods for Electric Vehicles, IEEE Access 6 (2017) 1832–1843.
[11] G. Rojas-Duenas, J.-R. Riba, M. Moreno-Eguilaz, Non-Linear Least Squares
In this paper, a methodology for black-box modeling of a 48 V –to- Optimization for Parametric Identification of DC-DC Converters, IEEE Trans. Power
Electron. 36 (1) (2021) 654–661.
12 V DC-DC bidirectional converter has been presented. The use of a [12] A. Frances, R. Asensi, O. Garcia, R. Prieto, J. Uceda, Modeling electronic power
deep learning algorithm enables the model to reproduce the behavior of converters in smart dc microgrids - An overview, IEEE Trans. Smart Grid 9 (6)
a nonlinear switching device, such as the bidirectional converter used in (2018) 6274–6287.
[13] V. Valdivia, A. Barrado, A. Lazaro, M. Sanz, D. Lopez Del Moral, C. Raga, Black-box
MHEVs. Based on the converter characteristics, a hyper-parameter behavioral modeling and identification of DC-DC converters with input current
tuning algorithm and a specific topology of the LSTM-NN have been control for fuel cell power conditioning, IEEE Trans. Ind. Electron. 61 (4) (2014)
discussed. An adequate experimental setup was needed to generate the 1891–1903.
[14] M. Al-Greer, M. Armstrong, M. Ahmeid, D. Giaouris, Advances on system
datasets to be used in the training process of the LSTM-NN. Both steady identification techniques for DC-DC switch mode power converter applications,
state and transient response data of the converter were identified with IEEE Trans. Power Electron. 34 (7) (2019) 6973–6990.
the same NN. The proposed model has a high accuracy when testing new [15] F. Alonge, F. D’Ippolito, T. Cangemi, Identification and robust control of dc/dc
converter Hammerstein model, IEEE Trans. Power Electron. 23 (6) (2008)
operating conditions different from those used during the training stage,
2990–3003.
while estimating the output voltage and current with low RMSE values [16] L. Arnedo, System Level Black-Box Models for DC-DC Converters, Virginia
and high coefficients of determination. The values of these indicators Polytechnic Institute and State University, 2008.
prove the robustness of the proposed model, which is able to replicate [17] G. Rojas-Duenas, J.-R. Riba, K. Kahalerras, M. Moreno-Eguilaz, A. Kadechkar,
A. Gomez-Pau, Black-Box Modelling of a DC-DC Buck Converter Based on a
the nonlinearities of the bidirectional converter under different oper­ Recurrent Neural Network, IEEE International Conference on Industrial
ating points. It has also been shown that the trained LSTM-NN out­ Technology (ICIT) 2020 (2020) 456–461.
performs three estimation methods commonly used to model power [18] P. Qashqai, K. Al-Haddad, R. Zgheib, in: in IECON Proceedings (Industrial Electronics
Conference), 2020, pp. 4697–4702.
converters in the technical literature. Finally, the proposed approach [19] M. Baumann, C. Weissinger, and H. G. Herzog, “System identification and
can also be applied for an accurate black-box modeling of many other modeling of an automotive bidirectional DC/DC converter,” in 2019 IEEE Vehicle
power electronics devices. Power and Propulsion Conference, VPPC 2019 - Proceedings, 2019.
[20] S.M. Ghamari, H. Mollaee, F. Khavari, Robust self-tuning regressive adaptive
controller design for a DC–DC BUCK converter, Meas. J. Int. Meas. Confed. 174
CRediT authorship contribution statement (2021), 109071.
[21] J. Bi, Y. Wang, S. Shao, Y. Cheng, Residual range estimation for battery electric
vehicle based on radial basis function neural network, Meas. J. Int. Meas. Confed.
Gabriel Rojas-Dueñas: Conceptualization, Software, Writing – 128 (2018) 197–203.
original draft, Validation. Jordi Roger Riba: Conceptualization, Re­ [22] W. Zhang, X. Li, X. Li, Deep learning-based prognostic approach for lithium-ion
sources, Writing – review & editing, Supervision. Manuel Moreno- batteries with adaptive time-series prediction and on-line validation, Meas. J. Int.
Meas. Confed. 164 (2020), 108052.
Eguilaz: Conceptualization, Supervision, Validation. [23] McKinsey & Company, “Mild hybrids—a multi-billion euro growth opportunity
alongside e-mobility?,” Aug. 2020.
[24] S.-R. Lee, B.-Y. Choi, J. Lee, and C.-Y. Won, “48 V to 12 V soft-switching
Declaration of Competing Interest interleaved buck and boost converter for mild-hybrid vehicle application,”
undefined, 2016.
[25] S. Belagoune, N. Bali, A. Bakdi, B. Baadji, K. Atif, Deep learning through LSTM
The authors declare that they have no known competing financial classification and regression for transmission line fault detection, diagnosis and
interests or personal relationships that could have appeared to influence location in large-scale multi-machine power systems, Measurement 177 (2021),
the work reported in this paper 109330.
[26] S. Hochreiter, J. Schmidhuber, Long Short-Term Memory, Neural Comput. 9 (8)
(1997) 1735–1780.
Acknowledgement [27] Y. Yu, X. Si, C. Hu, and J. Zhang, “A review of recurrent neural networks: Lstm cells
and network architectures,” Neural Computation, vol. 31, no. 7. MIT Press
Journals, pp. 1235–1270, 01-Jul-2019.
This work has been partially supported by the Generalitat de Cata­ [28] S. Wang, X. Wang, S. Wang, D. Wang, Bi-directional long short-term memory
lunya under the project 2017SGR0967, and the European Commission method based on attention mechanism and rolling update for short-term load
forecasting, Int. J. Electr. Power Energy Syst. 109 (2019) 470–479.
through the Clean Sky program, under the project 755332-AEMSIdFit.
[29] I. Sutskever, O. Vinyals, and Q. V Le, “Sequence to Sequence Learning with Neural
Networks,” 2014.
References [30] F. Karim, S. Majumdar, H. Darabi, S. Chen, LSTM Fully Convolutional Networks for
Time Series Classification, IEEE Access 6 (2017) 1662–1669.
[31] B. Shahriari, K. Swersky, Z. Wang, R. P. Adams, and N. De Freitas, “Taking the
[1] S. Saponara, P. Tisserand, P. Chassard, D.M. Ton, Design and measurement of
human out of the loop: A review of Bayesian optimization,” Proceedings of the
integrated converters for belt-driven starter-generator in 48 v micro/mild hybrid
IEEE, vol. 104, no. 1. Institute of Electrical and Electronics Engineers Inc., pp.
vehicles, IEEE Trans. Ind. Appl. 53 (4) (2017) 3936–3949.
148–175, 01-Jan-2016.
[2] S. Saponara, C.H.T. Lee, N.X. Wang, J.L. Kirtley, Electric Drives and Power
[32] H. Cheng, X. Ding, W. Zhou, R. Ding, A hybrid electricity price forecasting model
Chargers: Recent Solutions to Improve Performance and Energy Efficiency for
with Bayesian optimization for German energy exchange, Int. J. Electr. Power
Hybrid and Fully Electric Vehicles, IEEE Veh. Technol. Mag. 15 (1) (2020) 73–83.
Energy Syst. 110 (2019) 653–666.
[3] Z. Liu, S. Onori, A. Ivanco, Synthesis and Experimental Validation of Battery Aging
Test Profiles Based on Real-World Duty Cycles for 48-V Mild Hybrid Vehicles, IEEE
Trans. Veh. Technol. 66 (10) (2017) 8702–8709.

7
G. Rojas-Dueñas et al. Measurement 183 (2021) 109838

[33] D.P. Kingma, J.L. Ba, Adam: A method for stochastic optimization. 3rd [35] L. Wang, et al., Electromagnetic Transient Modeling and Simulation of Power
International Conference on Learning Representations, ICLR 2015 - Conference Converters Based on a Piecewise Generalized State Space Averaging Method, IEEE
Track Proceedings, 2015. Access 7 (2019) 12241–12251.
[34] A. Frances, R. Asensi, J. Uceda, Blackbox Polytopic Model with Dynamic Weighting
Functions for DC-DC Converters, IEEE Access 7 (2019) 160263–160273.