Energy Conversion and Management 296 (2023) 117668

Contents lists available at ScienceDirect

Energy Conversion and Management

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

Data-based modelling of proton exchange membrane fuel cell performance

and degradation dynamics
Adithya Legala a, Samaneh Shahgaldi a, b, Xianguo Li a, *
a
Department of Mechanical and Mechatronics Engineering, University of Waterloo, Waterloo, Canada
b
Hydrogen Research Institute (HRI), Université du Québec à Trois-Rivières (UQTR), Canada

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

Keywords: Proton exchange membrane fuel cell (PEMFC) is in the commercial adoption process for hard-to-decarbonize
Proton Exchange Membrane Fuel Cell (PEMFC) applications such as transport. However, its long-term durability, reliability, diagnostics, and performance
Artificial Neural Network (ANN) remain a critical challenge. In this study, the data-based modelling technique, artificial neural network (ANN), is
Accelerated Stress Test (AST)
adopted to comprehend performance degradation and thermal and pressure dynamics. Experimental data con­
Data-based models
Cyclic voltammetry
sisting of polarization and cyclic voltammetry data are acquired after every 6,000 accelerated stress test (AST)
Keras optimizers degradation cycles as stipulated by the US Department of Energy guidelines. Over 40,000 test cases were
considered for training and evaluating the neural network comprising data from 30,000 full-cell AST cycles. Fuel
cell operating temperatures, pressures, flow rates, and relative humidities are varied to capture the entire
spectrum of PEMFC performance characteristics. Parameters such as the number of AST cycles endured or
instantaneous catalyst electrode surface area, operating current density, reactant flow rate, relative humidity,
system temperature and pressure are considered as feature vectors (inputs) to predict cell output voltage,
reactant outlet pressure and temperature for both anode and cathode streams. Data preprocessing and batch
learning are implemented in Python to improve prediction accuracy and computational time; considering the
substantial nonlinear data sets, various Keras library optimizers and corresponding hyperparameters are inves­
tigated for better convergence. The trained neural network model is evaluated on 15,000 test cases resulting in
an R2 ≥ 0.995 for all the predicted variables, demonstrating the capability of data-based models to accurately
predict the nonlinear behavior of such electrochemical systems with minimal processing time, advocating for
their application in real-time system monitoring, controls and diagnostics.

open system, unlike a battery, can continuously generate electrical en­

ergy if reactants are fed into the system eliminating the device downtime
1. Introduction
for charging and avoiding the need for mammoth cell stacks [7,8].
Like any other electrochemical system, the performance and effi­
The impact of industrial and vehicular emissions on air quality and
ciency of the PEMFC depend on its operation and design, such as fuel cell
ecology has been recognized for a long time, especially during the Covid-
temperature, reactants supply pressures, the status of membrane hy­
19 pandemic [1,2]. It is imperative to decarbonize our economies for
dration (lambda), current density, the composition and structure of the
climate change mitigation and energy security [3,4]. The transport and
electrode (catalyst, ionomer, pore space), removal of water, flow dis­
energy sectors are at the forefront of undergoing this historical clean
tribution in cell structure and flow channels built on the bipolar plates
energy transition, and proton exchange membrane fuel cell (PEMFC) is
[9–12]. Further, its durability or degradation behavior is determined by
one of the most commercially anticipated technologies projected to lead
the mechanical and chemical stability of individual cell components
this transition [5,6]. PEMFC is a complex electrochemical device that
such as membrane, bipolar plates, and, most importantly, catalyst layers
converts the chemical energy of reactants directly into electrical energy
which consist of precious metals, support and ionomer, all of which are
through a series of oxidation and reduction reactions [7]. PEMFC’s high
affected by operating conditions [9–11,13]. The durability of these in­
efficiency, specific energy, and power density with zero emissions at the
dividual components affects the long-term performance, a fuel cell’s
point of use make it an ideal choice for heavy-duty transportation and
total useful life, and associated system diagnostics. A PEMFC operating
remote power applications [8]. PEMFC, being a thermodynamically

* Corresponding author.
E-mail address: Xianguo.Li@uwaterloo.ca (X. Li).

https://doi.org/10.1016/j.enconman.2023.117668
Received 23 May 2023; Received in revised form 12 September 2023; Accepted 13 September 2023
Available online 21 September 2023
0196-8904/© 2023 Elsevier Ltd. All rights reserved.
A. Legala et al. Energy Conversion and Management 296 (2023) 117668

degradation with an acceptable error but are limited to the ageing

Nomenclature mechanisms and assumptions considered [22,23]. Modern-day hard­
ware and computational limitations allow the deployment of ECM
Accelerated Stress Test AST models, which are computationally simple on the embedded controller.
Adaptive Moment Estimation Adam However, due to their complex physics-based differential equations, the
Artificial Neural Network ANN three-dimensional CFD models cannot be used on any real-time
Computational fluid dynamics CFD embedded systems. Modelling these individual degradation behaviors
electrode catalyst surface area ECSA through a physics-based or electro-chemistry-based model requires the
Equivalent Circuit Model ECM availability of detailed material properties, design attributes and a
Gas Diffusion Layer GDL comprehensive understanding of every physical phenomenon encoun­
Kernel Density Estimate KDE tered and their compounded effect of mutual interactions, which may
Learning Rate Lr not be possible at every instance leading to assumptions and default
membrane electrode assembly MEA limitations of the model capabilities. This complexity and lack of
Multi Input Multi Output MIMO transparency in fuel cell material properties make a case for adopting
Proton exchange membrane fuel cell PEMFC non-traditional techniques, such as artificial neural networks (ANN), to
Rectified linear unit ReLU resolve such complex issues [17,23–25].
Resilient Propagation Rprop Robust data-based models for PEMFC are in the early stage of
Stochastic gradient descent SGD development due to the inherent non-traditional modelling approach,
and limited availability of experimental or validated simulation data in
the public domain; physics-based models dominate. However, the rise of
computational power, data acquisition technology and the complexity
in a natural world environment is exposed to impurities such as sulfur, and interconnectivity of modern energy/propulsion systems are pushing
oxides of carbon, nitrogen and metals either through fuel or air, which towards a data-based approach. Chang et al [26] proposed a non-linear
can cause catalyst poisoning or deactivation, leading to chemical recurrent dynamic neural network to predict the remaining useful life of
degradation [9–11]. Similarly, thermal and humidity cycling under PEMFC by estimating the PEMFC voltage at various time intervals by
compression in the fuel cell causes hydrothermal ageing where the utilizing the PEMFC operational data of 150 to 650 h; however, this
membrane is subjected to excessive swelling and shrinking, which can approach involved the accumulation of large sets of experimental data
lead to catalyst leaching, sintering, deactivation, and loss of structural (operating the PEMFC for up to 1000 h) and also application of locally
membrane rigidity further leading to the formation of pinholes [13]. weighted regression to preprocess the data. Zuo [27] has used a long
Loss of pore space due to disruption of the gas diffusion layer (GDL) short-term memory network approach to predict the PEMFC degrada­
structure can lead to mass transport resistance, causing starvation and tion by estimating the cell voltage and this approach also consists of pre-
flooding inside the PEMFC electrodes. Constant voltage cycling conditioning the data by using a Savitzky-Golay filter to remove mea­
throughout the life of PEMFC can cause electrochemical-related ageing surement noise. Ma et al [28] applied a grid long short-term neural
leading to the degradation of metallic bipolar plates, catalyst layers and network to predict long-term voltage loss in the PEMFC where operating
membranes due to forming radicals. Depending on the operational current, voltage, and temperature are considered as feature vectors;
environment, application and calibration, these ageing mechanisms can however, utilizing this deep learning approach is computationally
simultaneously occur in a fuel cell, leading to various degradation be­ expensive just to predict the cell voltage in an on-board diagnostic setup.
haviors and characteristics [11]. Degradation in the PEMFC affects not Silva [29] has applied the adaptive neuro-fuzzy logic approach to
only the output voltage (or efficiency) of the cell but also its pressure and solving the problem of PEMFC prognostics focusing on real-time appli­
temperature-related dynamics due to the deactivation of the catalyst, cation to predict PEMFC voltage with great accuracy, however, this
change in pore space and membrane structural deterioration [14], approach does not consider the thermal and pressure dynamics. Simon
which in turn impact the performance degradation and reliability of the [30] has used the unique echo state network concept by replacing the
cells. hidden layers with neuron reservoirs and estimating the Hurst exponent
Additionally, unlike combustion engines, there are no significant in the ANN framework to predict cell voltage, but the algorithm was not
visual or audible indicators during operation for an electrochemical robust enough to comprehend the transitive behavior. Vichard [31] has
system like PEMFC that can be used as an identifier to detect degrada­ also used the echo state network-based neural network architecture by
tion or abnormalities before a catastrophic failure. Only the data from considering the total electric charge applied to the system as a relevant
internal stack electrical, thermal and pressure residual signatures can be input to accurately predict the cell voltage over 5000 h of PEMFC
used to identify degradation and malfunctions in the PEMFC; this is a experimental data, but results show that no less than 60% (3000 h)
significant safety challenge during the commercial operations of such training data is necessary for good accuracy. Pahon [32] demonstrated
systems [15–17]. Considering these PEMFC operational challenges, the application of signal processing techniques for fuel cell diagnostics
there is a definitive need for a PEMFC model that can accurately monitor and degradation by considering the relative entropy and output voltage;
the system response, predict the system performance, and estimate the however, the thermal and pressure dynamics of the system are not
degradation rate. The model must be able to predict output attributes considered limiting the possibility of detecting fuel cell flooding and hot
not only from a voltage perspective but also from a thermal and pressure spot formation. Chen [33] has combined the grey neural network model
dynamics perspective at various levels of cell degradation for real-time method with the particle swarm optimization technique to predict the
diagnostics and prognostics. degradation in the PEMFC by using a minimal amount of initial opera­
Traditionally, PEMFC uses the equivalent circuit models (ECM) to tional data but, only 50 h of data is considered for the validation, and
replicate the behavior of the fuel cell, predominantly for system controls only the cell voltage is predicted with nominal accuracy.
[18–21]; however, these models cannot comprehend the complete Most of the existing literature formulates the degradation behavior of
degradation or thermal or pressure dynamics. Computational fluid dy­ the PEMFC as a complex time series problem and needs hundreds of
namics (CFD) models for PEMFC are robust enough to predict the fluc­ hours of fuel cell operational data to capture the degradation behavior.
tuations in the thermal and pressure domain; however, due to the To the best of our knowledge, no existing data-based model can track
fundamental assumptions in the model, they cannot comprehend the accurately the degradation behavior along with temperature and pres­
complete degradation behavior over a long period of operations. sure dynamics of the PEMFC system that can be simultaneously used for
Physics-based electrochemical models can predict the state of health diagnostics and prognostics. Therefore, the objective of this study is to

2
A. Legala et al. Energy Conversion and Management 296 (2023) 117668

Fig. 1. Cross plot of PEMFC operational parameters [experimental data] (a) Fuel cell inputs − 1; (b) Fuel cell inputs – 2; (c) Fuel cell inputs – 3; (d) Fuel cell outputs.

tackle the above-mentioned gap in the literature by contributing the the concept of endured accelerated stress test (AST) cycles and
following: electrode catalyst surface area (ECSA).
3. The proposed method and training technique avoid the need for any
1. Development of a comprehensive multi-input multi-output data- additional signal conditioning and data filtering techniques,
based model that can predict the voltage degradation, temperature, demonstrating the model’s robust capability to predict voltage and
and pressure dynamics of the PEMFC across various operating pressure fluctuations.
conditions. 4. The proposed method is applicable for both steady conditions and
2. Formulating the PEMFC voltage degradation behavior as a regression during transition operations to predict temperature and voltage be­
problem, a proposed alternative to time series models by introducing haviors demonstrating the feasibility for on-board diagnostic
application.

3
A. Legala et al. Energy Conversion and Management 296 (2023) 117668

Fig. 1. (continued).

2. Experimental procedure compression to prevent any leakage of reactants during operation, thus
completing the PEMFC assembly. The entire experimental data is
First, a customized membrane electrode assembly (MEA) for the fuel generated using the Automated Greenlight G20 fuel cell test station. The
cell is manufactured in the lab using a commercially available 60% Pt/C initial performance (baseline) of the prepared fuel cell is evaluated by
catalyst. A homogenous catalyst ink is prepared with ethanol, deionized capturing the data from the polarization curve of the fuel cell at various
water, Aquivion as an ionomer and a total Pt (catalyst) loading of 0.5 operating conditions, such as at various stoichiometric ratios obtained
mg/cm2. The membrane is uniformly coated with the catalyst ink and by changing the flow rate of both hydrogen (0.1 to 4.5 slpm) and air
sandwiched between known porosity gas diffusion layers consisting of a (0.65 to 9 slpm), operating under various anode (0.5 to 270 kPa) and
microporous layer and a macroporous substrate, completing the seven cathode (0.5 to 230 kPa) pressures, at different fuel cell and reactant
layered MEA. The total active surface area of the MEA is around 45 cm2 temperatures (55 to 90 ◦ C), under various reactants relative humidity
and is encapsulated between the bipolar plates and sealed under conditions (20 to 100 % Rh). These PEMFC working conditions during

4
A. Legala et al. Energy Conversion and Management 296 (2023) 117668

Fig. 1. (continued).

the baseline evaluation are chosen to capture the data from the entire baseline evaluation and after degradation evaluation collected during
fuel cell operational spectrum where the sensitivities associated with the experiment are shown in Fig. 1. The data acquired during this
each input parameter to the fuel cell output can be modelled [34–36]. experimental procedure consists of an acceptable signal noise level and
After establishing the initial baseline, PEMFC is subjected to 30,000 is not subjected to any filtering or smoothening techniques to evaluate
accelerated stress test (AST) cycles to induce degradation of catalyst the algorithms’ robustness. Fig. 1 emphasizes the correlations between
layers by cycling the voltage through the fuel cell as per the U.S the various PEMFC experimental attributes using pair plots, which is a
Department of Energy guidelines [37]. The data from polarization combination of multiple scatter plots (here the nonlinearity of the data
curves and cyclic voltammetry are collected every 6000 cycles to can be visualized and analyzed) and kernel density estimate (KDE) plot
analyze the system performance degradation and electrode catalyst (diagonal plots illustrate the smoothed underlying data distribution).
surface area (ECSA). The cross plot of all the PEMFC’s operating con­ Scatter plots in Fig. 1 help identify if two corresponding variables have
ditions (system inputs) and polarization curve data (outputs) during the linear or nonlinear correlations (e.g., exponential, logarithmic, etc.) or if

5
A. Legala et al. Energy Conversion and Management 296 (2023) 117668

Fig. 1. (continued).

it is random (no correlation). Gap data analysis (gaps in operating endured or the instantaneous ECSA calculated by the cyclic voltammetry
ranges) from KDE plots of various parameters illustrated in Fig. 1 indi­ data as the input factor. Using the parameter of endured AST cycles/
cate that only reactant humidity and endplate temperature attributes ECSA as an input indirectly introduces the concept of fuel cell run time
have a clustered distribution, whereas the rest are dispersed across the into the model; here, a correlation between real-world fuel cell opera­
operational regime. Further, Fig. 1 also illustrates the variability of tional time and degradation associated with the number of AST cycles/
PEMFC outputs at different operating conditions and their respective ECSA is necessary, which varies on the type of application and system
nonlinearity and distribution with the operating current. calibration that can be determined experimentally. This approach
eliminates the need for complex time series-based recurrent neural
3. Data-based modeling networks or long short-term networks, reducing computational
complexity and processing time to predict the degradation level and
After considering nonlinearity in the data set and the initial literature associated corrections.
survey, the modelling aspect is formulated as a multi-input multi-output
regression problem and feed-forward artificial neural network (ANN)
architecture is chosen. The degradation aspect of the PEMFC is intro­ 3.1. Feed-forward ANN
duced into the model by either considering the number of AST cycles
Feed-forward ANN architecture is one of the most widely used data-

6
A. Legala et al. Energy Conversion and Management 296 (2023) 117668

Fig. 2. Schematic of the neural network [41].

3.2. Data preprocessing

Table 1
Feature vectors and prediction attributes for the present fuel cell ANN model.
Data preprocessing is necessary to transform the data in an organized
Feature Vectors (Inputs) Prediction attributes (Outputs) way at specific scales without altering the fundamental nature of the
ECSA/AST Cycles Cell Output Voltage data. Normalization is one of the most common preprocessing steps
Current Density Anode Pressure Out implemented on the input data before it is fed into a neural network, as
Anode Fuel Flow Cathode Pressure Out all the inputs and outputs correlate differently in nonlinearity and
Cathode Oxidant Flow Temp Anode Out
Anode Humidity Temp Cathode Out
sensitivity aspects. This preprocessing can help improve neural net­
Cathode Humidity work’s performance by making learning the weights of the synapses
Anode Operating Pressure (connections) between neurons easier. One of the most common ways to
Cathode Operating Pressure implement normalization on the data is the Min-Max Normalization
Temperature at Anode Bipolar Plate
method [42–45], where the data scales to a fixed range, usually between
Temperature at Cathode Bipolar Plate
Temperature Anode In 0 and 1. It is calculated by subtracting the minimum value from each
Temperature Cathode In data point and dividing the result by the range (maximum value minus
minimum value) as shown below:

based modelling techniques for supervised learning, which is inspired by XNormalized =

(X − XMinimum )
the human biological neural network consisting of layers of inter­ (XMaximum − XMinimum )
connected neurons [38–40]. A neuron in an ANN is the fundamental
In some cases, the decimal scaling method [42–45] can be imple­
processing unit, where it receives the information from other neurons,
mented where the data scales by multiplying each value by a power of
complies an output depending on embedded mathematical functions
10 and rounding to the nearest integer. This method can be helpful for
(also referred to as activation functions) and transmits output to the next
data with a wide range of values but a relatively small variance. In this
neuron. In a feed-forward ANN, the flow of information transmission is
case, the Min-Max normalization method is implemented as all the
only in one direction, i.e., from the input layers (input attributes are
PEMFC operational input parameters and output attributes are non-
introduced here) to subsequent hidden layers (which host inter­
negative physical entities with a smaller range.
connected neurons from multiple layers), and each node in the output
layer is connected to the nodes in the preceding hidden layer providing a
final output. A schematic of the neural network is shown in Fig. 2. The 3.3. PEMFC model feature vectors and prediction attributes
error backpropagation process calibrates the weights and biases asso­
ciated with each neuron based on the difference between the desired and In a neural network, the input dataset is typically organized into
targeted output. This error is then continuously propagated backwards multiple individual input arrays, also referred to as feature vector ar­
through each layer of the neurons to adjust their corresponding weights rays. These arrays of inputs represent the input variable characteristics
and biases until reaching the desired output that meets the specified at a particular sample or instance. For PEMFCs, feature vectors (input)
accuracy. are initially selected by considering the governing physics and chemical
The activation function embedded in the neurons introduces the kinetics involved and further screened based on the ability to measure or
nonlinear behavior to the model; the accuracy, training and computa­ estimate accurately and reliably. After deliberations, the following pa­
tional time requirements determine the number of hidden layers and rameters are chosen as the feature vectors (inputs) for the PEMFC ANN
associated neurons in each layer. Considering the nonlinearity of the model: the number of AST cycles endured or current electrode catalyst
data, training parameters such as learning rate, optimizers, and neuron surface area, operating current density, anode fuel flow rate, cathode
interconnection probabilities should be calibrated. All these factors, oxidant flow rate, anode reactant humidity, cathode oxidant humidity,
from the structure of the neural network to the training attributes, are anode operating pressure, cathode operating pressure, the temperature
considered hyperparameters tuned to achieve the required level of ac­ at anode bipolar plate, the temperature at cathode bipolar plate, the
curacy. Detailed information on ANN architecture, activation functions, anode inlet temperature, and cathode inlet temperature. The target
mathematical formulation and calibration is available elsewhere [41]. variables in an ANN, also called prediction variables, are the final out­
puts of the neural network, which can either be a classification decision

7
A. Legala et al. Energy Conversion and Management 296 (2023) 117668

Fig. 3. Schematic of the present fuel cell neural network.

Fig. 4. Graphical representation of Rectified linear unit function (red-colored curve) and derivative (blue-colored curve) [41]. (For interpretation of the references to
colour in this figure legend, the reader is referred to the web version of this article.)

or a regression value. For the PEMFC model, the prediction variables are model the PEMFC’s response in linear and nonlinear regions. The
chosen to capture the residual from electrical, thermal and pressure following hyperparameters and factors are considered during the initial
signatures. The cell output voltage, outlet pressure and temperature for ANN design and training phases.
both anode and cathode streams are chosen as prediction variables,
which can help diagnose the abnormal behaviour or estimate degrada­ 3.4.1. Activation function
tion performance. The feature vectors (inputs) and prediction variables Considering that the PEMFC polarization curve has both linear and
(outputs) for the present PEMFC model are listed in Table 1, and the nonlinear regions, the chosen activation function in the ANN should be
schematic of the neural network is shown in Fig. 3. able to comprehend both linear and nonlinear responses depending on
the feature vectors. The activation function introduces nonlinearity into
3.4. ANN design and hyperparameters the network and allows it to learn more complex relationships in the
data. Several different activation functions can be used in an ANN, each
The feed-forward ANN has multiple structural attributes and with unique properties and challenges such as Sigmoid, Tanh, and Swish
hyperparameters that must be optimized and calibrated to accurately [46–49]. However, most of the activation functions incorporate

8
A. Legala et al. Energy Conversion and Management 296 (2023) 117668

Fig. 5. Results of the neural network hyperparameter iteration: A; B; C; D.

9
A. Legala et al. Energy Conversion and Management 296 (2023) 117668

Fig. 5. (continued).

10
A. Legala et al. Energy Conversion and Management 296 (2023) 117668

Table 2
Neural network hyperparameter iteration details.
Iteration Hidden layers Neurons per layer Dropout Probability Optimizer Learning Rate Batch Size

A 2 50 0.25 SGD 0.025 32

B 3 150 0.5 SGD 0.0025 1000
C 3 500 0.5 Adam 0.0025 1000
D 2 150 0.75 Adam 0.005 500

derivative are shown in Fig. 4.

Table 3
Finalized neural network hyperparameters.
3.4.2. Hidden layers and nodes
Hyperparameters Values In the ANN, a hidden layer is defined as a layer of neurons that are
Inputs feature vectors 12 not directly connected to the feature vectors or prediction variables of
Output parameters 5 the network, whereas nodes are defined as the number of neurons in
Hidden layers 2
each hidden layer. ANN with less than two hidden layers is considered a
Nodes / Neurons 150
Activation function ReLU-1st,2nd layer; Linear – final layer shallow neural network, and a deep neural network has more than two
Loss function ‘MSE’ hidden layers. The number of hidden layers and nodes significantly
Optimizer Adam impacts the model’s ability to learn and generalize to new data and
Learning Rate (Lr) 0.005 requires experimentation to find the best configuration based on the
Beta_1 (decay factor) 0.9
Iterations / Epochs 1000
required accuracy, available computational power, model robustness,
Dropout probability 0.75 model training time and data set characteristics. Excessive hidden layers
Batch Size 500 and neurons can lead to overfitting of the model, which can be pre­
Training Time 5 Mins vented by regularization techniques, such as dropout and weight decay.
In this case, multiple iterations of hidden layers, neurons per hidden
layers, are simulated with a common activation function to evaluate the
Table 4 accuracy, training time and processing time. A few results of these it­
Sample points and fuel cell degradation levels. erations are discussed in the follow-up sections and illustrated in Fig. 5.
Sample Points Cell degradation level (Endured AST Cycles)
3.4.3. Optimizer, learning rate and epochs
1–132 6000
An optimizer is a mathematical algorithm that calibrates the weights
133––262 12,000
263–404 18,000
and biases of the connections between neurons in an artificial neural
405–534 24,000 network (ANN) to minimize the error between the predicted and true
535–668 30,000 values. The learning rate is the main hyperparameter of the optimizer
algorithm that needs to be calibrated, and this learning rate determines
the step size at which the optimizer modifies the weights. Too big of a
exponential operations that result in higher computational power and
learning rate may lead to skipping the global minima resulting in
processing time. Considering the computational efficiency, a non-
oscillation or divergence, and too small of a learning rate may lead to
exponential activation function Rectified linear unit (ReLU) [46–49] is
longer training times and processing power. Epoch represents a single
chosen for the present PEMFC modelling after multiple iterations, as it is
iteration of the training dataset, during which the weights and biases of
computationally simple and yields better results in both linear and
the connections between neurons are updated based on the error. The
nonlinear regions by addressing the vanishing gradient problem.
combination of learning rate and epochs should be balanced, as any
Graphical representation and behavior of the ReLU function and its
deficiency might result in either underfitting or overfitting the model.

Fig. 6. (a) Cell current at various sample points; (b) AST cycles at various sample points.

11
A. Legala et al. Energy Conversion and Management 296 (2023) 117668

Fig. 7. ANN predicted voltage degradation. (a) Voltage vs sample points; (b) Voltage vs cell current.

learning, the model is trained using a single instance of the data and
the weights and biases are updated immediately after each instance.
Batch learning is used when the dataset is large enough to be processed
in a reasonable amount of time and can be more computationally effi­
cient, leading to a better generalization performance, as it considers the
entire dataset when updating the weights. Whereas single-shot learning
is applied on small datasets, allowing the model to learn and adapt more
quickly to new data, it can be less accurate and more prone to over­
fitting, as it only considers the specific single instance to update the
weights and bias values. Considering the current data set is greater than
40,000 test cases, a Multi Input Multi Output (MIMO) regression batch
learning approach is implemented where the impact of different batch
sizes is investigated to compare the processing times and accuracy of the
model. A batch learning approach with a batch size of 500 to 5000
points resulted in a training time of under 10 min.
The prediction results of a few ANN model iterations and associated
hyperparameters are shown in Fig. 5 and tabulated in Table 2. As
illustrated in Fig. 5, the iteration process is initiated with Iteration A
(shallow neural network consisting of only two hidden layers and 50
neurons) with 0.25 dropout probability and associated hyperparameters
Fig. 8. Fuel cell reactant inlet temperature profile.
such as learning rate is tuned in an iterative process to achieve minimum
acceptable prediction error. The initial results from Iteration for pres­
Several different optimizers such as Stochastic gradient descent
sure attributes are excellent; however, voltage and temperature pre­
(SGD), Adaptive Moment Estimation (Adam), Resilient Propagation
diction accuracy could be better considering the normalized data. Based
(RProp), are available in Python’s Keras library for regression problems
on this observation, a second Iteration B is investigated (deep neural
[50–52]. The performance of both SGD and Adam optimizers are eval­
network consisting of only three hidden layers and 150 neurons) with
uated in this study: SGD updates the weights based on the gradient of the
0.5 dropout probability and similarly tuned with various hyper­
loss function with respect to associated weights. SGD has the advantage
parameters, and the performance of the neural network at the extreme
of being easy to implement and fast to compute, but it is sensitive to the
conditions is not adequate, in some cases, worse than Iteration A. Iter­
learning rate and may not always converge to a global minimum. Adam
ation C is built upon Iteration B by replacing the SGD optimiser with
optimizer on the other hand, combines the concept of SGD and mo­
Adam and increasing the neurons per layer which yielded a much better
mentum, where an exponentially decaying average of past gradients is
performance when compared to Iteration A and B. Iteration C despite
used to scale the learning rate, which can help the model converge more
demonstrating an excellent prediction accuracy is still a deep neural
robustly at a faster pace. For the Adam optimizer, both the learning rate
network with three hidden layers which takes a considerable amount of
and decay rate act as hyperparameters providing further flexibility in
training and processing time, to resolve this Iteration D with similar
calibration. The results of these iterations are discussed in the follow-up
Adam optimiser and increased dropout probability is configured on a
sections and illustrated in Fig. 5.
shallow neural network consisting of only two hidden layers and asso­
ciated hyperparameters are tuned. Iteration D performed relatively
3.4.4. Batch learning vs single-shot approach
similarly to Iteration C, with minimal training and processing time for
Batch and single-shot learning are two approaches to exposing the
the neural network.
data set to the ANN during training [53–55]. The difference between
After considering the iteration results, and computational efficiency,
both approaches is that in batch learning, the model is trained using a
iteration D had the best overall accuracy of R2 value greater than 0.995
fixed dataset, and the weights or biases of the connections between
and is selected as the final architecture and final hyperparameters of the
neurons are modified after processing the entire dataset. In single-shot

12
A. Legala et al. Energy Conversion and Management 296 (2023) 117668

Fig. 9. ANN predicted fuel cell temperature dynamics [ Red-Experimental Value; Blue–Predicted Value] (a) Anode reactant outlet temperature; (b) Cathode reactant
outlet temperature. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 10. Fuel cell reactant inlet pressure profile.

13
A. Legala et al. Energy Conversion and Management 296 (2023) 117668

Fig. 11. ANN predicted fuel cell pressure dynamics [ Red-Experimental Value; Blue–Predicted Value] (a) Anode reactant outlet pressure; (b) Cathode reactant outlet
pressure. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 12. ANN voltage prediction during transient conditions with measurement noise.

14
A. Legala et al. Energy Conversion and Management 296 (2023) 117668

Fig. 13. ANN pressure dynamics prediction during transient conditions with measurement noise.

ANN, as tabulated in Table 3. cycle is considered the new baseline. The polarization curve is evaluated
every 6000 cycles under the same conditions, and the model shows an
4. Results and discussion excellent correlation between the actual (true/experimental) and
model-predicted value. The sample data points of polarization curves
This section evaluates the PEMFC model performance after training with their actual and predicted voltage values at various current den­
by comparing the final prediction attributes to their actual experimental sities are shown in Fig. 7. As observed, the initial OCV at various levels of
values (true values). However, it must be noted that the experimental degradation is relatively unchanged however, the cell voltage at mid to
values recorded can have induced measurement noise due to instru­ higher current densities shows a considerable degradation in perfor­
mentation procedures and interference from another physiological mance and the model accurately captures the voltage degradation trend.
process. The performance of the PEMFC model to predict output voltage,
temperature, and pressure after various AST degradation cycles is dis­ 4.2. Prediction of PEMFC temperature dynamics
cussed in sections 4.1, 4.2, and 4.3, respectively. Sample points and
associated degradation conditions (AST cycles) and cell current are Similarly, the temperature of the outlet gases at both anode and
tabulated in Table 4 and represented in Fig. 6. The varying feature cathode is measured during the polarization curve evaluation at every
vector (input) data to the PEMFC model during the polarization curve 6000 cycles. The AST cycles had minimal effect on the temperature
evaluation are the operating cell current, and endured AST cycles, dynamics as expected due to the nature of the electrochemical ageing
whereas the relative humidity of reactants is maintained at 100%, the process, which does not impact the structural properties of the mem­
cell temperature is stabilized at 75 ◦ C, reactant inlet pressures are brane. Figs. 8 and 9 show that anode and cathode outlet temperatures
maintained at 35kPag, and the flow rate of hydrogen (H2) and air is set exhibit different patterns and trends of variations. Both exhibit tem­
at 4.5 slpm and 9 slpm, respectively. Section 4.4 discusses the PEMFC perature rises with increasing current density, but there is an observable
model effectiveness over transient operating conditions at widely varied dip on the anode side, suspected to be due to the back diffusion of water
temperatures, pressures, flow rates, humidities and current densities. from the cathode to the anode at higher operating currents reducing the
anode outlet temperature. The relative variability in temperature dip
4.1. Prediction of PEMFC voltage degradation between 6000 cycles and rest can be attributed to the change in the
water retention capacity of the membrane due to ageing, impacting the
For the first 6000 AST cycles, there was no observable (statistically water back diffusion process, thereby affecting the anode out reactant
significant) degradation in the cell performance; as a result, the 6000th temperature at high current densities [56]. The temperature change is

15
A. Legala et al. Energy Conversion and Management 296 (2023) 117668

Fig. 14. ANN prediction error histogram (a) Cell voltage; (b) Anode outlet pressure; (c) Cathode outlet pressure; (d) Anode outlet temperature; (e) Cathode outlet
temperature.

around 2 ◦ C to 4 ◦ C at normal operating conditions for a single cell; when deviate from normal behavior. It might be mentioned that the temper­
multiple cells and stacks are considered, this temperature change will be ature anomalies observed in Fig. 9(b) around the sample point of 400 at
larger, especially if a hotspot is encountered due to PEMFC pinhole the temperature of around 75.5 ◦ C are due to the transition in the
formation or due to any change in membrane ohmic resistance. operating current, that is, the operating current has been reduced after
Considering that the current model can predict similar minute temper­ the peak current condition, as this is also observable in both Figs. 6 and
ature patterns at various operating conditions, this approach can be 7.
scaled up for diagnostic purposes to predict hot spot formation and
membrane dehydration by recognizing the errors and patterns that

16
A. Legala et al. Energy Conversion and Management 296 (2023) 117668

4.3. Prediction of PEMFC pressure dynamics The computational time for a one-time step prediction of the voltage,
pressure, and temperature dynamics of the fuel cell using the trained
Similar to temperature dynamics, the PEMFC model can accurately ANN is 0.001997 s, less than two thousandth of a second, much faster
predict pressure drop across the fuel cell, the pressure drop across the than real time. This demonstrates the capability and capacity of the ANN
fuel cell changes with resistance to mass transport of reactants and model developed in this study to perform real-time control and online
products (all pressures are designated relative to gauge pressure). diagnosis.
However, during a constant reactant flowrate, the fluctuations in reac­
tant outlet pressures are minimal (only a few Pascals) unless there is a 5. Conclusions
significant blockage due to water; Figs. 10 and 11 illustrate only a
specific operating cycle with varying current and constant reactant flow In this study, a multi-input and multi-output data-based model for
rate (at various AST cycles), and the algorithm prediction performance proton exchange membrane fuel cell (PEMFC) trained on experimental
echoes the AST cycling, as seen in early figures. This is because the data using feed-forward neural network architecture is successfully
current data-based model is predominantly tuned to predict pressure developed to predict the voltage degradation, temperature, and pressure
fluctuations at transient conditions, for which it performs robustly with dynamics of fuel cell over various operating conditions. Twelve fuel cell
great accuracy as shown in the next subsection (section 4.4). As such, the operating parameters are considered as feature vectors (inputs) to
current model could be further tuned or optimized for this duty cycle of simultaneously predict five prediction attributes (fuel cell outputs) that
constant reactant flow rates; and the transient performance results at can be used to diagnose a fuel cell operation and degradation. An
different operating pressure ranges are shown in section 4.4. As ex­ experimental dataset consisting of over 40,000 test cases, including data
pected, the pressure drop across the cathode is higher, resulting in lower from various degradation levels is used for training and validation of the
cathode outlet pressures due to the high reactant flow rate. The AST neural network and is purposefully not subjected to any kind of data
cycles had minimal effect on the pressure dynamics as the degradation filtering or smoothening. The voltage degradation behavior is formu­
mechanism does not impact the gas diffusion layers, gaskets and bipolar lated as a regression problem, using the endured accelerated stress test
plate structure, which can be observed in the figures. A minute outlet (AST) cycles to input the neural network’s degradation factor. Multiple
pressure drop at high current conditions can be noted on the cathode iterations of model hyperparameters, optimizers and batch learning are
side, indicating minimal flooding. However, similar to the change in implemented to improve the accuracy and robustness of the neural
temperature, pressure drop across multiple cells and stacks will be a network. The model demonstrates excellent accuracy with R2 ≥ 0.995
magnitude higher and will increase during flooding at higher current for all the predicted variables upon normalized and pristine data during
densities. Considering the current PEMFC model’s performance level to the validation phase and on new data across multiple operating condi­
track pressure drop across a single cell, it can be scaled up to detect tions enabling the model to track voltage, temperature, and pressure
instances of flooding in PMEFC during abnormal operations. patterns, a majority within 5% error. The PEMFC neural network is able
to track the voltage, temperature, and pressure fluctuation in the pres­
4.4. Prediction of PEMFC transient dynamics and model robustness ence of degradation, measurement noise and transition conditions,
exhibiting great robustness and demonstrating the feasibility of such
Although fuel cells are designed primarily to operate under steady- models for onboard diagnostic purposes.
state conditions, the transient operating points are critical, and the
PEMFC model must show good correlations during these transient CRediT authorship contribution statement
conditions. The data acquisition system can also be subjected to mea­
surement noise during these transition processes due to the inherent Adithya Legala: Conceptualization, Methodology, Formal analysis,
limitations of sensors and data sampling intervals. The experimental Writing – original draft. Samaneh Shahgaldi: Data curation, Writing –
data has inherent measurement noise due to the limitation of current, review & editing. Xianguo Li: Methodology, Supervision, Project
voltage sensors’ precision (electrical interference) and the influence of administration, Funding acquisition, Conceptualization, Writing - re­
water condensation on thermocouples and pressure transducers. The view & editing.
data-based PEMFC model can not only predict the system outputs
accurately at different operating zones (high and low pressure, tem­
peratures and currents) but also has excellent robustness when dealing Declaration of Competing Interest
with sensor noise and unstable conditions, as shown in Figs. 12 and 13,
advocating its suitability and adoption for on-board diagnostics appli­ The authors declare that they have no known competing financial
cation. The algorithm also accurately predicts the pressure fluctuation in interests or personal relationships that could have appeared to influence
transient operation relatively better than the steady state, as illustrated the work reported in this paper.
in Fig. 13, across various pressure ranges. The overall performance of the
data-based PEMFC model’s prediction capability is evaluated across a Data availability
wide range of operating conditions, including steady state, transient,
low and high current densities, low and high reactant flow rates, pres­ Data will be made available on request.
sures and temperatures and the associated error statistics across 41,255
data points are represented using histograms as shown in Fig. 14. Most Acknowledgement
of the prediction attributes are within 5 % error, emphasizing the ac­
curacy and robustness of the algorithm, especially considering that there This work received financial support from Canadian Urban Transit
is no filtering of measurement data. Research and Innovation Consortium (CUTRIC) via Project Number
To evaluate the time taken for each model calculation for the 160028, and Natural Sciences and Engineering Research Council of
currently developed ANN model, a personal computer with the Canada (NSERC) via a Discovery Grant.
following specifications is used:
References
• Nvidia Graphic Card – RTX 3070 with 5467 MB device memory
• 12th Gen Intel(R) Core(TM) i9-12900, 2400 Mhz, 16 Core(s), 24 [1] le Quéré C, Jackson RB, Jones MW, Smith AJP, Abernethy S, Andrew RM, et al.
Temporary reduction in daily global CO2 emissions during the COVID-19 forced
Logical Processor(s) confinement. Nat Clim Chang 2020;10:647–53. https://doi.org/10.1038/s41558-
• Total computer system RAM of 31.7 GB 020-0797-x.

17
A. Legala et al. Energy Conversion and Management 296 (2023) 117668

[2] Yañez CC, Hopkins FM, Xu X, Tavares JF, Welch A, Czimczik CI. Reductions in Fuzzy Inference Systems. Int J Hydrogen Energy 2014;39:11128–44. https://doi.
California’s Urban Fossil Fuel CO2 Emissions During the COVID-19 Pandemic. AGU org/10.1016/J.IJHYDENE.2014.05.005.
Advances 2022;3. https://doi.org/10.1029/2022AV000732. [30] Morando S, Jemei S, Hissel D, Gouriveau R, Zerhouni N. Proton exchange
[3] Unfccc, ADOPTION OF THE PARIS AGREEMENT - Paris Agreement text English, n. membrane fuel cell ageing forecasting algorithm based on Echo State Network. Int
d. https://unfccc.int/sites/default/files/english_paris_agreement.pdf (accessed J Hydrogen Energy 2017;42:1472–80. https://doi.org/10.1016/j.
May 14, 2022). ijhydene.2016.05.286.
[4] Lempert R, Preston BL, Edmonds J, Clarke L, Wild T, Binsted M, et al. Climate [31] Vichard L, Harel F, Ravey A, Venet P, Hissel D. Degradation prediction of PEM fuel
Innovation 2019. cell based on artificial intelligence. Int J Hydrogen Energy 2020;45:14953–63.
[5] Canada. Natural Resources Canada, Hydrogen strategy for Canada : seizing the https://doi.org/10.1016/j.ijhydene.2020.03.209.
opportunities for hydrogen : a call to action., n.d. [32] Pahon E, Hissel D, Jemei S, Steiner NY. Signal-based diagnostic approach to
[6] A. Brooker, A. Birky, E. Reznicek, J. Gonder, C. Hunter, J. Lustbader, C. Zhang, L. enhance fuel cell durability. J Power Sources 2021;506. https://doi.org/10.1016/j.
Sittler, A. Yip, F. Yang, D.-Y. Lee, Vehicle Technologies and Hydrogen and Fuel Cell jpowsour.2021.230223.
Technologies Research and Development Programs Benefits Assessment Report for [33] Chen K, Laghrouche S, Djerdir A. Degradation prediction of proton exchange
2020, 2021. www.nrel.gov/publications. membrane fuel cell based on grey neural network model and particle swarm
[7] Li X. Principles of Fuel Cells, 1396. optimization. Energy Convers Manag 2019;195:810–8. https://doi.org/10.1016/j.
[8] Sazali N, Salleh WNW, Jamaludin AS, Razali MNM. New perspectives on fuel cell enconman.2019.05.045.
technology: A brief review. Membranes (Basel) 2020;10. https://doi.org/10.3390/ [34] Zhao J, Shahgaldi S, Ozden A, Alaefour IE, Li X, Hamdullahpur F. Effect of catalyst
membranes10050099. deposition on electrode structure, mass transport and performance of polymer
[9] Vichard L, Steiner NY, Zerhouni N, Hissel D. Hybrid fuel cell system degradation electrolyte membrane fuel cells. Appl Energy 2019;255. https://doi.org/10.1016/j.
modeling methods: a comprehensive review. J Power Sources 2021;506. https:// apenergy.2019.113802.
doi.org/10.1016/j.jpowsour.2021.230071. [35] Shahgaldi S, Alaefour I, Zhao J, Li X. Impact of ionomer in the catalyst layers on
[10] Wu J, Yuan XZ, Martin JJ, Wang H, Zhang J, Shen J, et al. A review of PEM fuel cell proton exchange membrane fuel cell performance under different reactant flows
durability: degradation mechanisms and mitigation strategies. J Power Sources and pressures. Fuel 2018;227:35–41. https://doi.org/10.1016/j.fuel.2018.04.076.
2008;184:104–19. https://doi.org/10.1016/j.jpowsour.2008.06.006. [36] Shahgaldi S, Ozden A, Li X, Hamdullahpur F. A scaled-up proton exchange
[11] Zhao J, Li X. A review of polymer electrolyte membrane fuel cell durability for membrane fuel cell with enhanced performance and durability. Appl Energy 2020;
vehicular applications: Degradation modes and experimental techniques. Energy 268. https://doi.org/10.1016/j.apenergy.2020.114956.
Convers Manag 2019;199:112022. https://doi.org/10.1016/j. [37] DOE CELL COMPONENT ACCELERATED STRESS TEST PROTOCOLS FOR PEM
enconman.2019.112022. FUEL CELLS, 2007. http://www1.eere.energy.gov/hydrogenandfuelcells/mypp/.
[12] Wang Y, Chen KS, Mishler J, Cho SC, Adroher XC. A review of polymer electrolyte [38] M. Dougherty, A REVIEW OF NEURAL NETWORKS APPLIED TO TRANSPORT,
membrane fuel cells: Technology, applications, and needs on fundamental 1995.
research. Appl Energy 2011;88:981–1007. https://doi.org/10.1016/j. [39] Hopfield JJ. Artificial Neural Networks. IEEE Circuits Devices Mag 1988;4:3–10.
apenergy.2010.09.030. https://doi.org/10.1109/101.8118.
[13] Jahnke T, Futter G, Latz A, Malkow T, Papakonstantinou G, Tsotridis G, et al. [40] Hammerstrom D. Working with neural networks. IEEE Spectr 1993;30:46–53.
Performance and degradation of Proton Exchange Membrane Fuel Cells: State of https://doi.org/10.1109/6.222230.
the art in modeling from atomistic to system scale. J Power Sources 2016;304: [41] Legala A, Zhao J, Li X. Machine learning modeling for proton exchange membrane
207–33. https://doi.org/10.1016/j.jpowsour.2015.11.041. fuel cell performance. Energy and AI 2022;10:100183. https://doi.org/10.1016/j.
[14] Whiteley M, Dunnett S, Jackson L. Failure mode and effect analysis, and fault tree egyai.2022.100183.
analysis of polymer electrolyte membrane fuel cells. Int J Hydrogen Energy 2016; [42] Eesa AS, Arabo WK. A Normalization Methods for Backpropagation: A Comparative
41:1187–202. https://doi.org/10.1016/j.ijhydene.2015.11.007. Study. Science Journal of University of Zakho 2017;5:319. https://doi.org/
[15] Forrai A, Funato H, Yanagita Y, Kato Y. Fuel-cell parameter estimation and 10.25271/2017.5.4.381.
diagnostics. IEEE Trans Energy Convers 2005;20:668–75. https://doi.org/ [43] Jayalakshmi T, Santhakumaran A. Statistical Normalization and Back
10.1109/TEC.2005.845516. Propagationfor Classification. International Journal of Computer Theory and
[16] G. Mousa, Fuel Cell Diagnostics using Electrochemical Impedance Spectroscopy, Engineering 2011:89–93. https://doi.org/10.7763/ijcte.2011.v3.288.
(2014). http://summit.sfu.ca/system/files/iritems1/14545/etd8628_GMousa.pdf. [44] Kim D. Normalization methods for input and output vectors in Backpropagation
[17] Petrone R, Zheng Z, Hissel D, Péra MC, Pianese C, Sorrentino M, et al. A review on neural networks. Int J Comput Math 1999;71:161–71. https://doi.org/10.1080/
model-based diagnosis methodologies for PEMFCs. Int J Hydrogen Energy 2013; 00207169908804800.
38:7077–91. https://doi.org/10.1016/j.ijhydene.2013.03.106. [45] Aksu G, Güzeller CO, Eser MT. The Effect of the Normalization Method Used in
[18] Asghari S, Mokmeli A, Samavati M. Study of PEM fuel cell performance by Different Sample Sizes on the Success of Artificial Neural Network Model,
electrochemical impedance spectroscopy, in. Int J Hydrogen Energy 2010: International Journal of Assessment Tools. Education 2019:170–92. https://doi.
9283–90. https://doi.org/10.1016/j.ijhydene.2010.03.069. org/10.21449/ijate.479404.
[19] Lazarou AAS, Karavotas K, Pyrgioti E. PEM Fuel Cell Equivalent Circuit Models. [46] S. Technická univerzita (Košice, IEEE Hungary Section, M. IEEE Systems, Institute
A Review, Power and Energy Systems (EuroPES 2008;2008):22–6. of Electrical and Electronics Engineers, SAMI 2020 : IEEE 18th World Symposium
[20] Page SC, Anbuky AH, Krumdieck SP, Brouwer J. Test method and equivalent on Applied Machine Intelligence and Informatics : proceedings : January 23-25,
circuit modeling of a PEM fuel cell in a passive state. IEEE Trans Energy Convers 2020, Herl’any, Slovakia, n.d.
2007;22:764–73. https://doi.org/10.1109/TEC.2007.895857. [47] B. Ding, H. Qian, J. Zhou, Activation functions and their characteristics in deep
[21] Lan T, Strunz K. Modeling of multi-physics transients in PEM fuel cells using neural networks, in: Proceedings of the 30th Chinese Control and Decision
equivalent circuits for consistent representation of electric, pneumatic, and thermal Conference, CCDC 2018, Institute of Electrical and Electronics Engineers Inc.,
quantities. Int J Electr Power Energy Syst 2020;119:105803. https://doi.org/ 2018: pp. 1836–1841. https://doi.org/10.1109/CCDC.2018.8407425.
10.1016/j.ijepes.2019.105803. [48] Apicella A, Donnarumma F, Isgrò F, Prevete R. A survey on modern trainable
[22] X. Zhang, P. Pisu, An Unscented Kalman Filter Based Approach for the Health- activation functions. Neural Netw 2021;138:14–32. https://doi.org/10.1016/j.
Monitoring and Prognostics of a Polymer Electrolyte Membrane Fuel Cell, n.d. neunet.2021.01.026.
[23] Zhao J, Li X, Shum C, McPhee J. A Review of physics-based and data-driven models [49] Szandała T. Review and comparison of commonly used activation functions for
for real-time control of polymer electrolyte membrane fuel cells. Energy and AI deep neural networks. In: Studies in Computational Intelligence. Springer; 2021.
2021;6:100114. https://doi.org/10.1016/j.egyai.2021.100114. p. 203–24. https://doi.org/10.1007/978-981-15-5495-7_11.
[24] S.-I. Ao, O. Castillo, C. Douglas, D.D. Feng, A. Korsunsky, International Association [50] Saleem MH, Potgieter J, Arif KM. Plant disease classification: A comparative
of Engineers, International MultiConference of Engineers and Computer Scientists : evaluation of convolutional neural networks and deep learning optimizers. Plants
IMECS 2016 : 16-18 March, 2016, the Royal Garden Hotel, Kowloon, Hong Kong, 2020;9:1–17. https://doi.org/10.3390/plants9101319.
n.d. [51] SCAD College of Engineering and Technology, Institute of Electrical and
[25] Mao L, Jackson L. Comparative study on prediction of fuel cell performance using Electronics Engineers, Proceedings of the International Conference on Trends in
machine learning approaches, Lecture Notes in Engineering and Computer. Science Electronics and Informatics (ICOEI 2019) : 23-25, April 2019, n.d.
2016;1:52–7. [52] S.K. Niranjan, K.L.S. Gogte Institute of Technology, Institute of Electrical and
[26] Chang TJ, Cheng SJ, Hsu CH, Miao JM, Chen SF. Prognostics for remaining useful Electronics Engineers. Bangalore Section., Institute of Electrical and Electronics
life estimation in proton exchange membrane fuel cell by dynamic recurrent neural Engineers, Proceedings of the International Conference on Computational
networks. Energy Rep 2022;8:9441–52. https://doi.org/10.1016/j. Techniques, Electronics and Mechanical Systems (CTEMS-2018) : 21 -23 December
egyr.2022.07.032. 2018, Belagavi, India, n.d.
[27] Zuo B, Cheng J, Zhang Z. Degradation prediction model for proton exchange [53] Runxuan Z. Efficient sequential and batch learning artificial neural network
membrane fuel cells based on long short-term memory neural network and methods for classification problems; 2005.
Savitzky-Golay filter. Int J Hydrogen Energy 2021;46:15928–37. https://doi.org/ [54] Lin J, Zhou DX. Online learning algorithms can converge comparably fast as batch
10.1016/j.ijhydene.2021.02.069. learning. IEEE Trans Neural Netw Learn Syst 2018;29:2367–78. https://doi.org/
[28] Ma R, Yang T, Breaz E, Li Z, Briois P, Gao F. Data-driven proton exchange 10.1109/TNNLS.2017.2677970.
membrane fuel cell degradation predication through deep learning method. Appl [55] Qin SZ, te Su H, McAvoy TJ. Comparison of four neural net learning methods for
Energy 2018;231:102–15. https://doi.org/10.1016/j.apenergy.2018.09.111. dynamic system identification. IEEE Trans Neural Netw 1992;3:122–30. https://
[29] Silva RE, Gouriveau R, Jemeï S, Hissel D, Boulon L, Agbossou K, et al. Proton doi.org/10.1109/72.105425.
exchange membrane fuel cell degradation prediction based on Adaptive Neuro- [56] Collette FM, Lorentz C, Gebel G, Thominette F. Hygrothermal aging of Nafion®.
J Memb Sci 2009;330:21–9. https://doi.org/10.1016/j.memsci.2008.11.048.

18