i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 4 5 ( 2 0 2 0 ) 1 4 9 5 3 e1 4 9 6 3

Available online at www.sciencedirect.com

ScienceDirect

journal homepage: www.elsevier.com/locate/he

Degradation prediction of PEM fuel cell based on

artificial intelligence

L. Vichard a,b,*, F. Harel b,d, A. Ravey a,b, P. Venet b,c, D. Hissel a,b
a
FEMTO-ST Institute, Univ. Bourgogne Franche-Comt
e, UTBM, CNRS, Belfort, France
b
FCLAB, CNRS, Belfort, France
c
Univ Lyon, Universit
e Claude Bernard Lyon 1, Ecole Centrale de Lyon, INSA Lyon, CNRS, Ampere, Villeurbanne,
France
d
AME-Eco7, Univ Gustave Eiffel, IFSTTAR, Univ Lyon, Lyon, France

highlights graphical abstract

Fuel cell 5000 h durability test

under transportation conditions.

Performance loss lower than 10%
after 5000 operating hours.

Degradation model based on Echo
state neural network.

Accurate performance prediction
over 2000 h.

article info abstract

Article history: In the last years, Proton Exchange Membrane Fuel Cells (PEMFC) became a promising en-
Received 28 January 2020 ergy converter for both transportation and stationary applications. However, durability of
Received in revised form fuel cells still needs to be improved to achieve a widespread deployment. Degradation
24 March 2020 mechanisms and aging laws are not yet fully understood. Therefore, long-term durability
Accepted 25 March 2020 tests are necessary to get more information. Moreover, degradation models are requested
Available online 16 April 2020 to estimate the remaining useful life of the system and take adequate corrective actions to
optimize durability and availability. This paper presents in a first part the results of a long-
Keywords: term durability test performed on an open cathode fuel cell system operated during 5000 h
Fuel cell system under specific operating conditions including start/stop and variable ambient temperature.
Fuel cell aging Performance evolution and degradation mechanisms are then analyzed to understand
Long term durability test influence of operating conditions and how to extend the durability. In a second part of the
paper, the results are used to build a degradation model based on echo state neural

* Corresponding author. FEMTO-ST Institute, Univ. Bourgogne Franche-Comte
, UTBM, CNRS, Belfort, France.
E-mail address: loic.vichard@utbm.fr (L. Vichard).
https://doi.org/10.1016/j.ijhydene.2020.03.209
0360-3199/© 2020 Hydrogen Energy Publications LLC. Published by Elsevier Ltd. All rights reserved.
14954 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 4 5 ( 2 0 2 0 ) 1 4 9 5 3 e1 4 9 6 3

Degradation model network in order to predict the performance evolution. Results of the degradation pre-
Echo state network diction are very promising as the normalized root mean square error remains very low
with a prediction time over 2000 h.
© 2020 Hydrogen Energy Publications LLC. Published by Elsevier Ltd. All rights reserved.

system for 5000 h (DOE target) under operating conditions

Introduction close to transportation applications.
In a first part, the prognostic discipline will be introduced
Today’s electric vehicles are mainly powered by lithium-ion and existing tools in the literature will be presented. In a
batteries and there is a long way before they become domi- second part, the experimental context and the long-term
nant in the global automotive market. Autonomy of electric durability tests will be detailed. The fuel cell system, the
vehicles is still too weak to compete with traditional gas ve- cycling profile and the chosen health indicators will be high-
hicles [1]. A solution to improve the autonomy is to hybridize lighted, then results will be drawn. Performance evolution of
the powertrain by including not only batteries but also alter- the system will be observed and influence of operating con-
native electrochemical devices [2]. Proton Exchange Mem- ditions and operating time on degradation and performance
brane Fuel Cell (PEMFC) is regarded to have a great potential will be analyzed. Based on these analyses, a method to model
due to the merits of high energy density, high conversion ef- performance evolution using echo state neural networks, with
ficiency, low operating temperature and zero gas emission [3]. respect to operating time and according to operating condi-
However, their durability is still a limitation and needs to be tions will be detailed. The structure of the network, its pa-
improved to achieve a worldwide commercialization [2]. In- rameters, the learning scheme and chosen inputs will be
tegrated in a vehicle, a fuel cell system is subjected to rough presented. Finally, predictions provided by the model will be
operating conditions [4,5], such as load variations [6], air im- highlighted and discussed. The obtained degradation model
purities, temperature and humidity cycling, cold ambient can be used to estimate remaining useful life, to anticipate
temperatures [7,8], vibrations and shocks which accelerate maintenance operation or to optimize energy management in
their degradation [4,9]. In 2010, the US Department Of Energy order to extend the system durability.
(DOE) fixed a durability target at 5000 h for transportation
applications whereas current durability is around 4000 h. For
several years many researches have carried out to improve Prognostic
this durability. Researches focus on improving the system
design and the choice of materials, and on reducing the causes Background
of degradation by implementing optimal supervision and en-
ergy management strategies. The objective of the prognostic discipline is to estimate the
As part of a continuous durability improving approach, remaining useful life of a system in order to anticipate
prognostic methods allow to estimate the end of life of the maintenance operations and to adapt energy management.
components and then deciding adequate actions at the right The concept of prognostic discipline is highlighted in the
time in order to extend possibly the durability [10,11]. Fig. 1. The purpose is to use historical data describing the
Based on historical data describing the performance evo- evolution of a State Of Health (SOH) indicator with respect to
lution, prognostic methods model the degradation which then the operating time and according to operating conditions, to
allows predicting the state of health of the considered system. build a degradation model. Once the model is built, it can be
In the case of Proton Exchange Membrane Fuel Cell (PEMFC), used with new inputs to predict the SOH indicator i.e. the
different approaches have been proposed in the literature performance evolution, with respect to the operating time and
which can be classified according to three categories: model- according to operating conditions.
based methods, data-driven methods and hybrid methods
[11,12].
To build a suitable degradation model, a good under-
standing of the system performance evolution and the
degradation mechanisms occurring within the system is pri-
mordial. Nowadays, degradation laws and mechanisms of fuel
cell are not fully understood. Then, experimental studies with
a system operating on test benches or integrated in actual
applications are important to get information about these
degradation mechanisms. In the literature, there is a lack a
long-term durability tests which are costly and time
demanding. For this purpose, this study carried out a long-
term durability test on a complete open cathode fuel cell
Fig. 1 e Prognostic concept.
 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 4 5 ( 2 0 2 0 ) 1 4 9 5 3 e1 4 9 6 3 14955

When the predicted SOH indicator value reaches a be very difficult. These methods are therefore rather dedi-
threshold of failure defined by the user, the remaining useful cated to simple systems whose physic laws and degradation
life can be quantified as the difference between the predicted mechanisms are well understood.
time when threshold is reached and the present time.
Generally the end of life is defined as 10% of performance loss, Data-driven methods
versus beginning of life and for automotive fuel cell stacks [10] Data driven methods are considered as black box approaches
but the end of life can be also defined as the inability to meet to realize a prognostic. They do not require system knowledge
user specifications [13]. and physical laws to start a prognostic. Measured data and
historical data are used to learn the dynamic behavior of the
Methods system. Therefore, data-driven prognostic methods are
particularly relevant for complex systems whose degradation
Several methods to model the performance evolution can be laws are not well understood. Napoli et al. [20] used a classical
found in the literature. These methods can be classified ac- neural network to predict stack voltage and cathode temper-
cording to three categories: model-based, data-driven and ature of a 5 kW fuel cell. Results showed that the proposed
hybrids [10e12]. method can reproduce the impact of different stoichiometric
ratio on the voltage under different operating conditions.
Model-based methods Ibrahim et al. [21] used a discrete wavelet transform to predict
Model-based methods are based on analytical models to the power of a fuel cell. Results showed that the method is
reproduce the dynamic behavior of the studied system ac- satisfying as the prediction error is less than 3%. Silva et al.
cording to the degradation mechanisms. The main advantage [22] propose a data-driven method to predict the output
of the model-based methods is that their do not require a large voltage reduction of a fuel cell caused by degradation during
amount of data. Lechartier et al. [14] proposed a degradation nominal operating conditions. The proposed method is based
model of a fuel cell composed of a static part and a dynamic on an Adaptative Neuro-Fuzzy Inference Systems (ANFIS) and
part that are independent. The static part is based on equa- use as an input the measure output voltage of the fuel cell.
tions describing physical phenomena and Butler-Volmer law. Results show that the proposed method is suitable to predict
The dynamic part is based on an electrical equivalency of the voltage loss. Liu et al. [23] proposed a data-based method
physical phenomena. Parameters of the models are updated using Long Short-term memory (LSTM) recurrent neural
based on characterizations: polarization curves and electro- network (RNN) to estimate the remaining useful life of a
chemical impedance spectroscopy. Bressel et al. [15] used an PEMFC. Results are verified with experimental aging data and
extended Kalman filter to estimate the state of health of a fuel show that the prediction accuracy of the proposed method is
cell and to predict its degradation evolution. Bressel et al. [16] 99.23% and the root mean square error is 0.003. Zhu et al. [24]
also proposed a model based method using the energetic presented a data driven method for inferring the remaining
macroscopic representation with time varying parameters to useful life of a fuel cell. They proposed a method based on the
forecast the fuel cell remaining useful life for a given power use of a Gaussian process state space model which allows to
reference. Lee et al. [17] proposed a method for estimating the estimate the evolution of the latent states and the future
state and remaining useful life of a single PEMFC cell through behavior of the fuel cell voltage. Ma et al. [25] proposed a
an equivalent impedance model based on electrochemical prognostic data-driven method based on the use of Grid Long
impedance spectroscopy (EIS). The author uses charge trans- Short Term Memory (G-LSTM). The proposed method was
fer resistance and double layer capacitance as major param- experimentally validated, and results indicate that the pro-
eters for judging degradation. EIS is a very suitable tool to get posed Grid long short-term memory network can predict the
information about state of health of a fuel cell system but for fuel cell degradation in a precise way. Morando et al. [26] used
now dedicated to laboratory conditions. Kimotho et al. [18] an Echo State neural Network (ESN) for fuel cell prognostic
proposed a prognostic approach based on adaptive particle purpose. They model and predict the fuel cell voltage evolu-
filter. The proposed method lies in the introduction of a self- tion with a mean average percentage error lower than 5%.
healing factor after each characterization and the adaptation Moreover, the study revealed that the computational re-
of the model parameters to fit the degradation evolution. Re- quirements for ESN are very low. Hua et al. [27] also used the
sults show that prognostic method is reliable with majority of ESN as a prognostic method. They also ensure that ESN
the predictions falling within 5% error. Zhang et al. [19] pro- structure has demonstrated better performances especially in
posed an approach based on two physical models. The first reducing the computational requirements. The study com-
one receives a signal directly observable and related to the pares the single-input structure with a multiple-input struc-
stack voltage. The second one is fed periodically by mea- ture. Results show that the multiple-inputs has a better
surements from physical characterizations of the stack. performance under both static and dynamic operating
Characterizations give good information about the SOH evo- conditions.
lution. Then, the prognostic procedure is implemented using
Particle Filtering (PF). Hybrid methods
For a complex system whose degradation mechanisms are Hybrid approaches combine data-driven approaches and
influenced by many factors, inferring an analytical model can model-based approaches, cumulating advantages of both
14956 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 4 5 ( 2 0 2 0 ) 1 4 9 5 3 e1 4 9 6 3

approaches. Jouin et al. [28] proposed a hybrid prognostic

method which aims at predicting the power losses of a fuel
cell stack running under constant operating conditions and
constant current. The proposed method is based on a new
empirical modeling for power aging and joint particle filters
framework. When enough data are available, the prediction of
the behavior is promising compared to experimentation.
Remaining useful life can be estimated over 500 h with an
error lower than 5%. Cheng et al. [29] proposed a hybrid
prognostic method for a fuel cell. The method is based on a
data-driven approach: Least Square Support Vector Machine
(LSSVM), combined with a model-based approach: regularized
particle filter. The results showed that the proposed hybrid
method combines both advantages of data-driven and model-
based approaches by providing a higher accuracy of the pre-
dicted remaining useful life. Liu et al. [30] propose an hybrid
Fig. 2 e A vehicle of the Mobypost fleet.
method to predict the degradation evolution and estimate the
remaining useful life of a PEMFC under different current loads.
In a first part the machine learning based on an Adaptative
Neuro-Fuzzy Inference System (ANFIS) tool is used to predict
the long-term degradation evolution. In a second part the
remaining useful life is estimated by using a semi-empirical
degradation model based on Unscented Kalman Filter algo-
rithm (UKF). Zhou et al. [31] proposed an hybrid method to
estimate the remaining useful life of a PEMFC. The approach is
based on the combination of a physical aging model and an
autoregressive and moving average (ARMA) model. The pre-
diction accuracy and robustness are experimentally demon-
strated with degradation tests performed on two types of
PEMFC stack.
Hybrid methods are tough to be developed. Indeed,
combining data-driven approaches and model-based ap- Fig. 3 e Mobypost vehicle architecture.
proaches allow to cumulate advantages of both approaches
but the drawbacks are also cumulated. Therefore, these
vehicles is performed in real time all along the driving cycles
methods require high amount of data and a good under-
which has led to create a significant database.
standing of the physical laws.
Consequently, among the three categories presented pre-
PEMFC in fuel cell hydrogen electric vehicles
viously, a data-driven method has been preferred in this study
and specially a method based on the use of a neural network.
In such vehicles, an open-cathode PEMFC system is integrated
Degradation laws and mechanisms of fuel cells are still not
as a range extender in the powertrain architecture highlighted
fully mastered and understood then physical models are very
in Fig. 3. The PEMFC system powers the DC bus with a con-
complex to build. In addition, in this study the operating
stant current to recharge the battery pack during stop phases
conditions are changed all along the experimentation with a
variable ambient temperature. Moreover, the studied system
is a commercial system then information about components
characteristics are not available. Therefore, a neural network
which is independent of the degradation laws is particularly
relevant and suitable for this purpose as shown in Jemei et al.
[32] and will be preferred in this study.

Experimental context

Mobypost project

The European project Mobypost aimed at designing and

developing a fleet of ten fuel cell hydrogen electric vehicles
(FCHEV) [15,16]. Theses compact and lightweight vehicles,
illustrated in Fig. 2, are dedicated to postal delivery applica-
tions. A real-time monitoring of all the components of the Fig. 4 e Fuel cell system.
 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 4 5 ( 2 0 2 0 ) 1 4 9 5 3 e1 4 9 6 3 14957

or to support it during driving phases. The battery pack pro- ambient air remains identical all along the experiments. To
vides the current dynamics required for the propulsion of the ensure the environmental conditions inside the climatic
vehicle through two in-wheels permanent magnet synchro- chamber and constantly renew the air, an air circulation was
nous machines set up with their associated inverters. set up as it can be seen in Fig. 5.

PEMFC system Cycling profile

This PEMFC system, illustrated in Fig. 4, is a commercial The purpose of the durability test was here to analyze in lab
AIRCELL™ system from H2SYS company. The system in- the system performance evolution when considering the
tegrates an open cathode fuel cell stack, an air supply line, a same operating conditions than on the actual FC vehicle. A
hydrogen supply line, a cooling management system, hu- postal delivery lasts about four hours during which the fuel
midification system, electronics to control electrical fluxes cell system provides a quite constant current to the DC bus.
and a controller to supervise actions and safety functions. Its Based on this application, the Fig. 6 highlights the cycling
nominal power is 1 kW and the stack is composed of 28 cells. profile used to age the system on the laboratory test bench
The open cathode stack with forced-convection and dead-end under the same conditions that it would operate in a range
anode is air cooled by fans and self-humidified by short- extender architecture [33]. In order to respect average postal
circuits. The nominal operating point is 50 A, 62  C (stack delivery duration, the system provides a constant current at
temperature) and the ambient operating temperature range its nominal operating point (50 A, 60  C) for four hours. This
specified by the manufacturer is þ5  Ce35  C. operating phase is followed by two hours of rest by turning off
Once the system is started up, it is self-powered and the system. This cycle of six hours was repeated continuously
communicates through a CAN bus network its physical to allow accelerated analysis of performance degradation.
operating data such as fuel cell temperature, stack current Considering this cycling profile, 5000 operating hours equals
and voltage of the twenty-eight cells. to 1250 startups/shutdowns and is equivalent to approxi-
mately 4 years of vehicle service with a daily postal delivery
cycle.
Long term durability test Otherwise, integrated in a vehicle means the system
operates under a variable ambient temperature and this
Test bench ambient temperature mainly depends on the period of the
year. To consider these variations, the system was integrated
The Fig. 5 proposes an illustration of the autonomous test into a climatic chamber. The ambient temperature value was
bench used for the durability test. This test bench is composed changed per period to approach seasonal variations. The
of a climatic chamber, an active load, a fluids control and an ambient temperature evolution is presented in the Table 1.
electronic control unit dedicated to control and communica-
tions with the fuel cell system. The test bench also integrates State of health indicator
its own sensors.
During experiments, the fuel cell system is placed in a 650L To follow performance evolution and quantify degradation, a
ADF type climate chamber (20  C to þ100  C, ±0.5  C) from SOH indicator has to be defined. The SOH of a fuel cell system
BIA climatic company. The climatic chamber controls the refers to the quantification of the performance potential
ambient temperature. It is worth noting that in the experi- compared with a nominal reference. The SOH can be quanti-
mental room, the air relative hygrometry is controlled to be fied by the loss of power that the system can provide for a
around 50% at 21  C. This point signifies that water content in specific current value. In the case of this study, the system
output current remains unchanged then the system health
evolution can be observed with the system output voltage.

Voltage evolution

To follow system output voltage evolution during all the

experimentation, the operating conditions must be identical.
To respect this rule, once a day, the average output voltage is
computed during a duration of 1000 s. In this period, the fuel
cell temperature is stabilized around 60  C and nominal
operating conditions are assured. The results of this durability
test are presented on the Fig. 7. The system output voltage is
highlighted with respect to the operating time and the
ambient temperature.
These results lead to the following observations:

Stage 1: during the first 1000 h the ambient temperature

was set up at 20  C. It can be observed that the system
Fig. 5 e Test bench.
14958 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 4 5 ( 2 0 2 0 ) 1 4 9 5 3 e1 4 9 6 3

Fig. 6 e Cycling profile.

output voltage decreases constantly overtime due to nat- temperature seems to lead to a recovery of the system
ural degradation of the components. performance with long term phenomena.

Stage 2: the ambient temperature was then elevated and
Stage 4: the ambient temperature was reduced again and
set up at 30  C for the next 1000 h. As soon as the ambient set up at 7  C. Once more, as soon as the temperature was
temperature was elevated, the slope of the system output reduced, a slight variation of the operating point can be
voltage drop is increased. Higher ambient temperatures noticed. Contrary to the previous stage, here the system
seem to accelerate degradation of the system. output voltage clearly increases during the entire stage.

Stage 3: the ambient temperature was reduced and set up This verify previous observations, reducing the ambient
at 20  C. As soon as the ambient temperature was reduced temperature seems to lead to a recovery of the system
a voltage increase can be observed: the operating point performance with long term phenomena. Moreover, low
seems to be changed. Moreover, the system output voltage ambient temperature seems to reduce the system degra-
evolution remains quite constant during the 400 h. dation significantly.
Degradation is no more noticeable. Reducing the ambient
Stage 5: the ambient temperature was elevated and set up
at 20  C for the next 1000 h. The degradation slope returned
negative as the system output voltage decreases with a
Table 1 e Ambient temperature evolution. constant slope quite closed to the stage 1.
Stage 1 Operating time Ambient temperature
Stage 6: the ambient temperature was reduced and set up
duration 10  C for the next 1000 h. Once more, as soon as the
1 1000 h 20  C ambient temperature was reduced a variation of the
2 1000 h 30  C operating point can be noticed: direct elevation of the
3 400 h 20  C voltage. Moreover, all along the stage, the system output
4 600 h 7 C voltage increases constantly. Degradation slope became
5 1000 h 20  C positive as in stage 4 what verifies previous observations.
6 1000 h 10  C
 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 4 5 ( 2 0 2 0 ) 1 4 9 5 3 e1 4 9 6 3 14959

Fig. 7 e Evolution of system output voltage all along the experimentation.

To synthesize, the operating time leads to an irreversible higher the ambient temperature, the lower the water content
degradation on the system as the global performance trend is in the fuel cell membranes. Then, reducing the ambient
to decrease. Otherwise, an impact of the ambient temperature temperature leads to re-moisture of the membranes which
on the performance and the degradation can be observed with can explains the long-term performance recoveries observed
short-term and long-term phenomena. Firstly, short-term during the experimentation.
variations of the operating point are directly linked to the To conclude on this durability test, the system has oper-
ambient temperature value. Secondly, long-term phenomena ated for 5000 h under variable ambient temperature condi-
acting on the degradation are depending on the ambient tions and numerous start/stop cycles and the performance
temperature: the higher the ambient temperature, the higher degradation is lower than 10%. Global efficiency is higher than
the degradation, the lower the ambient temperature the lower 49% and degradation rate is 11mV/cell/h.
the degradation. Moreover, reducing ambient temperature This kind of system: open cathode, air-cooled and self-
seems to lead to a recovery of the system performance as the humidified seems to be a very promising energy converter
degradation slope (performance evolution) becomes positive. integrated in a low power range extender architecture for
transportation applications. Moreover, experimentation car-
Analysis ried out in this study showed that better humidification con-
ditions would lead to a significant reduction of the
Previous section showed that ambient temperature variations degradation so a significant durability improvement.
lead to a variation of the operating point which is linked to the Based on this durability test and the presented aging data
net output power supplied by the system. This means that the purpose of this study is to propose a method to model the
ambient temperature directly affects the net output power degradation according to the operating time and ambient
supplied by the system. The system output current remains temperature variations which have shown a significant in-
equal to 50 A all along the experimentation, what implies that fluence on the system performance.
the power required by ancillaries inside the system varies
according to the ambient temperature. This variation of the
power required by the ancillaries can be explained by the air- Degradation model
cooling regulation to maintain the fuel cell stack temperature
at 60  C. The higher the ambient temperature, the higher the Several structures of neural network have been presented in
air flow rate to cool the fuel cell stack, the higher the power the literature. Recurrent Neural Networks (RNN) are particu-
required by the fans. larly suitable to model non-linear temporal signals. Morando
This link between ambient temperature and cooling air et al. [35] proposed a comparative study of different neural
flow can explain long-term phenomena acting on system network architectures used for fuel cell prognostic purpose.
degradation. In this system, the humidification strategy re- They showed that RNN and more specially Echo State Net-
mains unchanged whatever the ambient temperature. works (ESN) provide very good results and very interesting
Therefore, if the air flow increases for higher ambient tem- computing times. Consequently, ESN with multiple inputs will
peratures then the evacuation of the water created by elec- be preferred as a tool to model the performance evolution of
trochemical reactions will be accelerated. In other words, the the studied system.
14960 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 4 5 ( 2 0 2 0 ) 1 4 9 5 3 e1 4 9 6 3

with the desired spectral radius, the reservoir matrix is

divided by its spectral radius and then multiplied by the
desired spectral radius.

Learning scheme
The learning scheme of an ESN is governed by the following
steps. In a first time, the reservoir update x~ðnÞ is computed as
in the equation (1).


Fig. 8 e ESN architecture. x~ðnÞ ¼ f Win :uðnÞ þ Wres :xðn  1Þ (1)

with x~ðnÞ the reservoir update, uðnÞ the ESN input and
Echo state network xðn  1Þthe output value of the reservoir according to the
previous state. Wres is normalized by dividing it by its
Background & structure spectral radius value and then multiplied by the desired
Echo state network was introduced by Jaeger et al. [36] at the spectral radius value. Based on the previous equation, the
beginning of the 2000s. This type of recurrent neural network reservoir output can be express as following:
reproduces more faithfully the functioning of the human
xðnÞ ¼ ð1  aÞ:xðn  1Þ þ a:x~ðnÞ (2)
brain as the hidden layers are replaced by a reservoir of neu-
rons. The structure is highlighted in the Fig. 8. with a the leaking rate.
with: Then can be determined the output of the ESN yðnÞ from
the previous equations:
- Nres the number of neurons in the reservoir.
- K the number of inputs. yðnÞ ¼ f ðWout :xðnÞÞ (3)
- L the number of outputs. where f is a nonlinear function as a sigmoid, tangent, etc. The
- Win the input weight matrix. purpose of the learning scheme is to optimize the output
- Wout the output weight matrix. matrix Wout in order to estimate the output yðnÞ of the ESN as
- Wres the reservoir weight matrix. close as possible to the real value ytarget by minimizing the
quadratic error (4).
The input weight matrix and the weight matrix of the
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
reservoir are created randomly. Concerning the optimization u
u1 X T
 2
εðy; ytarget Þ ¼ t
target
of the weights, only the output weight matrix is optimized by a yi ðnÞ  yi ðnÞ (4)
T n¼1
multi linear regression. Therefore, computing times to build
the model are much lower compared to feedforward neural with T the number of samples.
network or again Long Short-Term Memory neural networks
(LSTM). Model inputs

ESN parameters The purpose of a degradation model is to reproduce and pre-

Some parameters configure an ESN. They can be listed as dict the performance evolution of a system according to the
following: operating conditions. Results of the durability tests high-
lighted that the performance and the degradation of the fuel
- The number Nres of neurons in the reservoir. Theoretically, cell system show relationships with the ambient temperature
a high number of neurons leads to a better accuracy of the and the operating time. In our case, the operating time vari-
model. However, it is important to pay attention to not able is meaningful as the system operates at nominal oper-
define a too important number of neurons since compu- ating point during the entirety of the durability test. In the
tation times are very closely related. case of the current profile is different, the operating time
- The leaking rate a: Between 0 and 1, the leaking rate can be variable would not be relevant as aging image. To tackle this
defined as the ability to forget the previous states. In other issue, the total electric charge (Ah) supplied by the system
words, the leaking rate refers to the influence of previous would be much more meaningful as input variable referring to
states on the current state. the aging. Accordingly, in this study ambient temperature and
- The spectral radius of the reservoir weight matrix, between operating time are defined as inputs of the proposed model.
0 and 1, refers to the effective time constant of the ESN. Moreover, a fuel cell system is a dynamic system. It implies
Higher spectral radius leads to a slower decay of impulse that the output system voltage at previous states is linked to
response. The spectral radius value also determines the the output voltage at present state through a non-linear dif-
amount of nonlinear interaction of input components ferential relationship. Consequently, for the proposed model,
through the time. This parameter corresponds to the output voltage at previous state is defined as an input.
maximum eigenvalues of the reservoir matrix. The reser- Obviously, adding more input variables leads to a better
voir matrix is created randomly then its spectral radius model generality. In our case, the load current is constant then
value is also random. In order to obtain a reservoir matrix it is not relevant as an input variable. In the case of a different
 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 4 5 ( 2 0 2 0 ) 1 4 9 5 3 e1 4 9 6 3 14961

current profile, considering this current as an input variable Firstly, the learning rate was set at 60%. The learning rate is
would have been required. Moreover, in our case and as defined as the ratio between the quantity of data used for the
already explained, the water content in the ambient air is training and the total quantity of data in the database. With a
unchanged all along the test, then the relative humidity would learning rate set at 60%, the model has been built and trained
not be relevant as an input variable. Facing another situation, with the entire ambient temperature range, then performance
considering the relative humidity as an input variable would recoveries observed previously are learnt. Results are high-
have been relevant. lighted in the Fig. 9. It can be observed that the model is able to
Nevertheless, adding more input variables would lead to predict the voltage evolution with a very good reliability
higher computational requirements and higher equipment whatever the ambient temperature. The Normalized Root
requirements such as sensors, wiring, data logging, etc. This is Mean Square Error (NRMSE) is equal to 0.098 with a prediction
a reason why a good compromise between model accuracy over 2000 h.
and measured variables must be reached. The learning rate value is really important, and users must
To summarize, the proposed model includes three inputs: be very carefully according to their application. In a second
ambient temperature, operating time and output voltage at time, the learning rate was set at 33%. A learning rate set at
previous state. It is worth noting that once the model is built, 33% refers to an operating duration about 1600 h and an
the output voltage at previous state becomes the predicted ambient temperature range between 20  C and 30  C used to
output voltage at previous state. build the model. The results of the performance prediction are
presented in the Fig. 10.
Results In this figure, it can be observed that the results are not
satisfying. The prediction, is close to the reality from 1600 h to
The model has been developed with the Anaconda™ envi- 2000h where the ambient temperature is 30  C. Once the
ronment and the used programming language is python ambient temperature is reduced to 20  C (stage 3) the pre-
which is suitable for machine learning purpose. The leaking dicted degradation is attenuated in the image of the stage 1
rate is set at 0.5 and the spectral radius is set at 0.5. These but the performance recovery is not reproduced by the model.
parameters were identified empirically. The computing time The same thing can be observed for stages 4 and 6 when
to build the model is about two seconds. The computer used to ambient temperature is 7  C and 10  C. The predicted voltage
build the model runs with Windows 10 and integrates a CPU evolution presents a negative slope, lower than the other
Intel core I7-6800k 3.40 GHz, a GPU NVIDIA TITAN X and 64 stages but not positive as expected. To summarize with a
giga of memory.

Fig. 9 e Results with learning rate ¼ 60%. Fig. 10 e Results with learning rate ¼ 33%.
14962 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 4 5 ( 2 0 2 0 ) 1 4 9 5 3 e1 4 9 6 3

evolution of a complete open cathode fuel cell system under

transportation conditions: start/stop and variable ambient
temperature. Then, results were analyzed in order to under-
stand performance evolution. Analysis showed that the
ambient temperature has an influence on the performance
and the degradation of the system with short and long term
phenomena. Low ambient temperature leads to better hu-
midification conditions and significantly reduces the degra-
dation rate. This system seems very reliable for the
application as the average degradation rate is 11mV/cell/h after
5000 operating hours and the performance loss is 9% versus
the beginning of life. It must be underlined that these exper-
imental results are better than the durability target perfor-
mances set by the US DOE for automotive applications.
Finally, measured data were used to build a degradation
model based on echo state neural network. Results show that
Fig. 11 e Results with learning rate ¼ 20%.
the proposed prognostic tool is very suitable as the degrada-
tion prediction is very accurate with a prediction time over
2000 h whatever the ambient temperature and computing
learning rate set at 33% the performance recovery phenomena
times about 2 s.
are not reproduced as the model did not learn them.
In a same way, the Fig. 11 highlights a prediction with a
learning rate set at 20% which refers to the first 1000 operating
Perspectives
hours. In this case, only the first stage has been used to train
and build the model. This signifies that the model has been
Analysis of the aging data and the understanding of the
trained with an ambient temperature value always equal to
degradation mechanisms and causes following the long-term
20  C. Therefore, in this case the model does not know that the
durability test will be used to develop an intelligent energy
ambient temperature affects the system performance and
management and humidification strategy in order to extend
degradation, not even higher ambient temperature (30  C).
the fuel cell system durability.
This is the reason why the performance evolution is quite
In addition, the proposed prognostic tools can be inte-
constant as a continuity of the first 1000 h. In the Fig. 10,
grated in future systems in order to estimate in real time the
despite the performance recovery phenomena are not
remaining useful life and to anticipate the maintenance op-
considered, the ambient temperature effects are considered in
erations. It can be also used to adapt energy management
some ways as the performance evolution slopes varies with
strategy to optimize the system durability.
the ambient temperature. Contrary, as shown in Fig. 11, with a
learning rate of 20% the ambient temperature value does not
affect prediction at all. references
Depending on the application and the operating condi-
tions, the influence of the learning rate on the model reli-
ability can be very significative. In the case of the threshold of [1] Saw LH, Ye Y, Tay AAO. Integration issues of lithium-ion
failure is defined at 17.5 V then with a learning rate set at 33% battery into electric vehicles battery pack. J Clean Prod
the estimated end of life of the system would be about 3000 h. 2016;113:1032e45.
[2] Lü X, Qu Y, Wang Y, Qin C, Liu G. A comprehensive review on
With a learning rate set at 60% then the estimated end of life
hybrid power system for PEMFC-HEV: issues and strategies.
seems to be higher than 6000 h.
Energy Convers Manag 2018;171:1273e91.
To conclude on this prognostic tool, the reliability is very [3] Vadiee A, Yaghoubi M, Sardella M, Farjam P. Energy analysis
promising as the NRMSE is equal to 0.098 with a prediction of fuel cell system for commercial greenhouse application e
over 2000 h and 3000 h of learning under variable operating a feasibility study. Energy Convers Manag 2015;89:925e32.
conditions. It could be integrated online to model perfor- [4] Pei P, Chen H. Main factors affecting the lifetime of Proton
mance evolution and estimate in real time the remaining Exchange Membrane fuel cells in vehicle applications: a
review. Appl Energy 2014;125:60e75.
useful life of the system based on historical measured data.
[5] Zhao J, Li X. A review of polymer electrolyte membrane fuel
The higher the number of measured data, the higher the ac- cell durability for vehicular applications: degradation modes
curacy of the model. and experimental techniques. Energy Convers Manag
2019;199:112022.
[6] Wahdame B, Candusso D, François X, Harel F, Pe
ra M-C,
Conclusion Hissel D, Kauffmann J-M. Comparison between two PEM fuel
cell durability tests performed at constant current and under
solicitations linked to transport mission profile. Int J
This study presents a method to model the degradation of a
Hydrogen Energy 2007;32:4523e36.
fuel cell system in order to estimate its remaining useful [7] Oszcipok M, Riemann D, Kronenwett U, Kreideweis M,
lifetime. Firstly, a long-term durability test of 5000 h was Zedda M. Statistic analysis of operational influences on the
presented. This test was carried out to observe performance
 i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 4 5 ( 2 0 2 0 ) 1 4 9 5 3 e1 4 9 6 3 14963

cold start behaviour of PEM fuel cells. J Power Sources [21] Ibrahim M, Steiner NY, Jemei S, Hissel D. Wavelet-based
2005;145:407e15. approach for online fuel cell remaining useful lifetime
[8] Jia L, Tan Z, Kang M, Zhang Z. Experimental investigation on prediction. IEEE Trans Ind Electron 2016;63:5057e68.
dynamic characteristics of proton exchange membrane fuel [22] Silva RE, Gouriveau R, Jemeı̈ S, Hissel D, Boulon L,
cells at subzero temperatures. Int J Hydrogen Energy Agbossou K, Yousfi Steiner N. Proton exchange membrane
2014;39:11120e7. fuel cell degradation prediction based on Adaptive Neuro-
[9] Zhang T, Wang P, Chen H, Pei P. A review of automotive Fuzzy Inference Systems. Int J Hydrogen Energy
proton exchange membrane fuel cell degradation under 2014;39:11128e44.
start-stop operating condition. Appl Energy 2018;223:249e62. [23] Liu H, Chen J, Hou M, Shao Z, Su H. Data-based short-term
[10] Jouin M, Gouriveau R, Hissel D, Pe
ra M-C, Zerhouni N. prognostics for proton exchange membrane fuel cells. Int J
Prognostics and health management of PEMFC e state of the Hydrogen Energy 2017;42:20791e808.
art and remaining challenges. Int J Hydrogen Energy [24] Zhu L, Chen J. Prognostics of PEM fuel cells based on
2013;38:15307e17. Gaussian process state space models. Energy 2018;149:63e73.
[11] Sutharssan T, Montalvao D, Chen YK, Wang W-C, Pisac C, [25] Ma R, Yang T, Breaz E, Li Z, Briois P, Gao F. Data-driven
Elemara H. A review on prognostics and health monitoring of proton exchange membrane fuel cell degradation
proton exchange membrane fuel cell. Renew Sustain Energy predication through deep learning method. Appl Energy
Rev 2017;75:440e50. 2018;231:102e15.
[12] Jouin M, Bressel M, Morando S, Gouriveau R, Hissel D, Pe
ra M- [26] Morando S, Jemei S, Hissel D, Gouriveau R, Zerhouni N.
C, Zerhouni N, Jemei S, Hilairet M, Ould Bouamama B. Proton exchange membrane fuel cell ageing forecasting
Estimating the end-of-life of PEM fuel cells: guidelines and algorithm based on Echo State Network. Int J Hydrogen
metrics. Appl Energy 2016;177:87e97. Energy 2017;42:1472e80.
[13] Lechartier E. Contribution au prognostic de pile a [27] Hua Z, Zheng Z, Pe
ra M-C, Gao F. Remaining useful life
combustible PEMFC base
sur modele semi-analytique. PhD prediction of PEMFC systems based on the multi-input echo
thesis. Universite
de Franche Comte
; 2017. state network. Appl Energy 2020;265:114e791.
[14]
ra M-C, Gouriveau R, Hissel D,
Lechartier E, Laffly E, Pe [28] Jouin M, Gouriveau R, Hissel D, Pe
ra M-C, Zerhouni N.
Zerhouni N. Proton exchange membrane fuel cell behavioral Prognostics of PEM fuel cell in a particle filtering framework.
model suitable for prognostics. Int J Hydrogen Energy Int J Hydrogen Energy 2014;39:481e94.
2015;40:8384e97. [29] Cheng Y, Zerhouni N, Lu C. A hybrid remaining useful life
[15] Bressel M, Hilairet M, Hissel D, Ould Bouamama B. Extended prognostic method for proton exchange membrane fuel cell.
Kalman filter for prognostic of proton exchange membrane Int J Hydrogen Energy 2018;43:12314e27.
fuel cell. Appl Energy 2016;164:220e7. [30] Liu H, Chen J, Hissel D, Su H. Remaining useful life
[16] Bressel M, Hilairet M, Hissel D, Ould Bouamama B. Model- estimation for proton exchange membrane fuel cells using a
based aging tolerant control with power loss prediction of hybrid method. Appl Energy 2019;237:910e9.
Proton Exchange Membrane Fuel Cell. Int J Hydrogen Energy [31] Zhou D, Al-Durra A, Zhang K, Ravey A, Gao F. Online
2020;45:11242e54. remaining useful lifetime prediction of proton exchange
[17] Lee J, Lee J-H, Choi W, Park K-W, Sun H-Y, Oh J-H. membrane fuel cells using a novel robust methodology. J
Development of a method to estimate the lifespan of proton Power Sources 2018;399:314e28.
exchange membrane fuel cell using electrochemical
seaux de Neurones to cite this version:
[32] Jemeı̈ S. Par Re
impedance spectroscopy. J Power Sources 2010;195:6001e7. Mode
lisation d ’ une pile a combustible de type PEM par
[18] Kimotho JK, Meyer T, Sextro W. PEM fuel cell prognostics
seaux de neurones. PhD thesis. Universite
re
de Technologie
using particle filter with model parameter adaptation. In: de Belfort Montbe
liard; 2004.
2014 Int. Conf. Progn. Heal. Manag. PHM 2014; 2015. p. 1e6. [33] Ravey A, Faivre S, Higel C, Harel F, Djerdir A. Energy
[19] Zhang D, Baraldi P, Cadet C, Yousfi-Steiner N, Be
renguer C, management of fuel cell electric vehicle with hydrid tanks.
Zio E. An ensemble of models for integrating dependent In: IECON Proc. Industrial Electron. Conf.; 2014. p. 3962e7.
sources of information for the prognosis of the remaining [35] Morando S. Contribution au Pronostic de Dure
e de Vie d’une
useful life of Proton Exchange Membrane Fuel Cells. Mech  Combustible a
Pile a  Membrane Echangeuse de Protons. PhD
Syst Signal Process 2019;124:479e501. thesis. Universite
de Franche Comte
; 2015.
[20] Napoli G, Ferraro M, Sergi F, Brunaccini G, Antonucci V. Data [36] Jaeger H. The “echo state” approach to analysing and training
driven models for a PEM fuel cell stack performance recurrent neural networks-with an erratum note. Bonn Ger
prediction. Int J Hydrogen Energy 2013;38:11628e38. Ger Natl Res Cent Inf Technol GMD Tech Rep 2001;148.