This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS 1

Decoupling Control of Fuel Cell Air Supply System

Based on Data-Driven Feedforward and Adaptive
Generalized Supertwisting Algorithm
Lin Chen , Jinfa Liu , Shihong Ding , Senior Member, IEEE, Jing Zhao , Member, IEEE, Jinwu Gao ,
and Hong Chen , Fellow, IEEE

Abstract— Decoupling control of the air supply system is is responsible for delivering fresh air to the cathode at the
crucial for enhancing the performance and prolonging the service appropriate flow rate and pressure.
life of proton exchange membrane (PEM) fuel cells. However, The air flow rate has a significant impact on the power
the strong coupling and nonlinearity inherent in the system pose
significant challenges. Current decoupling techniques typically
response of the fuel cell and the aging of the PEM, while the
rely on model knowledge and commonly overlook the avoidance cathode pressure is closely associated with mass transfer and
of compressor surge, which motivates our work with a twofold PEM stress. Low air flow rates can lead to oxygen starvation,
contribution. We first design a data-driven feedforward (DDF) resulting in reduced fuel cell output performance and potential
and propose a feasible domain constraint (FDC) to avoid surge. damage to the PEM [3]. Additionally, higher cathode pressure
Subsequently, an adaptive generalized supertwisting algorithm can enhance oxygen diffusion through the GDL, thereby
(AGSTA) is presented that eliminates the residual tracking
errors of the DDF. Furthermore, its gradient descent principle accelerating the electrochemical reaction [4]. It also promotes
and stability are demonstrated. The proposed method has been consistent distribution and reduces local hypoxia [5]. Further-
validated on an air supply system test bench and a hardware- more, maintaining a stable cathode pressure can minimize
in-the-loop (HiL) platform carrying a fuel cell electric vehicle stress variations on the PEM [6]. Accordingly, precise control
(FCEV) model. The results indicate that our approach is more of air supply systems is crucial for enhancing fuel cell power
advantageous in terms of tracking accuracy, response speed,
overshoot suppression and computational cost. response and mitigating stack aging. However, the strong
coupling and nonlinearity of air supply systems pose great
Index Terms— Decoupling control, adaptive control, extremum challenges for air supply control [7].
seeking, generalized supertwisting algorithm, air supply system,
proton exchange membrane fuel cell.
N OMENCLATURE
I. I NTRODUCTION Wca,d /Wca Reference/actual value of cathode inlet flow

A S AN emerging energy technology, PEM fuel cells

have been widely applied in various fields due to the
advantages of high efficiency and power density, low operating
Wca,o
W O2 ,r e
rate
Cathode outlet flow rate
Oxygen flow rate reacted by fuel cell stack
temperature, as well as virtually zero emissions [1]. The Pca,d /Pca Reference/actual value of cathode pressure
PEM fuel cell consists primarily of a stack, an air supply N /Nact Control command/actual value of compressor
system, a hydrogen supply system, a cooling system, and a speed
power management system [2], where the air supply system N f /Nb DDF/feedback component of N
θ/θact Control command/actual value throttle angle
Received 5 August 2024; revised 31 October 2024 and 10 December 2024; θ f /θb DDF/feedback component of θ
accepted 2 January 2025. This work was supported in part by the Major Sci-
ence and Technology Project of Jilin Province under Grant 20220301010GX
R Universal gas constant
and in part by the International Scientific and Technological Cooperation Vca Volume of cathode supply manifold
under Grant 20240402071GH. This article was recommended by Associate Ist /Tst Stack current/temperature
Editor Y. Tang. (Corresponding author: Jinwu Gao.)
Lin Chen, Jinfa Liu, and Jinwu Gao are with the State Key Laboratory
Mair /M O2 Molar mass of air/oxygen
of Automotive Simulation and Control, Department of Control Science and γ Specific heat ratio of air
Engineering, Jilin University, Changchun 130022, China (e-mail: chenlin21@ C D.tr Throttle discharge coefficient
mails.jlu.edu.cn; [email protected]; [email protected]). A T,tr Maximum throttle opening area
Shihong Ding is with the School of Electrical and Information Engineering,
Jiangsu University, Zhenjiang 212013, China (e-mail: [email protected]). n ce Number of cells in stack
Jing Zhao is with the School of Mechanical Engineering and F Faraday constant
Automation, Northeastern University, Shenyang 110819, China (e-mail: ai /ωi Perturbation amplitude/frequency of ES
[email protected]).
Hong Chen is with the Department of Control Science and Engineer-
µi Update gain of ES
ing, Tongji University, Shanghai 200092, China (e-mail: chenhong2019@ T Iterative period for cost function of ES
tongji.edu.cn). q1 /q2 Weighting of tracking error of Wca /Pca
Digital Object Identifier 10.1109/TCSI.2025.3526144
h Cut-off frequency of washout filter
1549-8328 © 2024 IEEE. All rights reserved, including rights for text and data mining, and training of artificial intelligence
and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: GUANGZHOU UNIVERSITY. Downloaded on February 02,2025 at 10:16:59 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS

βi Parameter optimized by ES DMD-based and DDC-based technologies. The essence of

βi∗ Local optimal value of βi DMD lies in model identification and decoupling matrix,
β̂i Estimate of unknow βi∗ which transforms the air supply system from a MIMO cou-
pling system to two SISO systems. Liu et al. [7] developed
A BBREVIATIONS a feedforward decoupling controller using DMD to quickly
ADRC Active disturbance rejection control respond to variable load requirements while avoiding oscil-
AGSTA Adaptive GSTA lations and steady-state errors in system response. Zhao et
DC Direct current al. [21] proposed a dynamic decoupling strategy based on
DDC Data-driven control fuzzy logic using DMD, which realizes the rapid adjustment
DDF Data-driven feedforward of flow rate and pressure stability. Yuan et al. [22] combined
DMD Diagonal matrix decoupling with DMD, SMC, ADRC meanwhile and proposed a cas-
ECU Electronic control unit cade control scheme, flow rate and pressure tracking in the
ES Extremum seeking external and internal loops, respectively. Additionally, Zeng et
ET Error tolerance al. [23] developed a DMD-based decoupling scheme which
ES-AGSTA ES-based adaptive GSTA effectively reduces the adverse effects of coupling. Qiu et
FCEV Fuel cell electric vehicle al. [24] presented an optimization of DMD to avoid adverse
FDC Feasible domain constraint conditions and improve performance. These studies contribute
GDL Gas diffusion layer significantly to the decoupling of air supply systems, but DMD
GSTA Generalized supertwisting algorithm requires model identification and decoupling matrices, which
LQR Linear quadratic regulator come with enormous computational effort.
MIMO Multiple-input-multiple-output Recently, DDC has also been utilized in air supply systems.
NMPC Nonlinear model predictive control Zhang et al. [25] developed an adaptive SMC for the flow rate
PEM Proton exchange membrane based on a novel data-driven sliding surface, and employed
PID Proportional-integral-derivative a backstep method to regulate the pressure. Furthermore,
SISO Single-input-single-output Li et al. [11] derived a control-oriented model by com-
SMC Sliding mode control bining data-driven and system mechanism, and proposed an
observer-based decoupling control strategy. The above studies
Considerable efforts have been dedicated to air supply employ the DDC and leverage partial model prior knowledge,
systems, leading to the development of several control-oriented which have driven advances in the development of DDC within
models. Pukrushpan et al. [8] created a high-precision model air supply systems. However, their control performance is
for a fuel cell system, which served as the foundation for still partially constrained by the model prior knowledge, and
numerous reduced-order models [9], [10], [11]. Additionally, deteriorates as environmental conditions alter or the system
Saleh et al. [12] developed a simplified low-power fuel cell ages.
model based on commercial 1 kW fuel cells. The aforemen- Although the aforementioned decoupling approaches
tioned models play a significant role in advancing air supply achieve decent performance, they all rely on model knowledge,
control. either identified or known a priori. Accordingly, it is worth-
Based on control-oriented models, several strategies have while to consider whether an alternative scheme exists that
been developed to regulate the flow rate of air supply systems. frees the control performance from the constraints of model
Chen et al. [13] utilized feedback linearization, which is knowledge. Moreover, inappropriate [Wca,d , Pca,d ] cause com-
effective when the model is dynamically matched to the pressor surges, resulting in significant mechanical vibrations
system. Pukrushpan et al. [8] developed a LQR controller and potential damage to the mechanical components. However,
based on linearized model, which possesses limited abil- there has been rare research on how to prevent compressor
ity to handle strong nonlinearities. The nonlinear triple-step surges in existing control strategies.
method [14], [15] has also been employed, demonstrating Motivated to address the aforementioned issues, this arti-
superior performance in tracking the desired flow rate and cle proposes a decoupling method that does not rely on
suppressing interference. SMC was commonly used to prevent model knowledge. Additionally, constraints on [Wca,d , Pca,d ]
oxygen starvation during sudden load changes and had the are considered to prevent compressor surge. The presented
ability to counter disturbances and nonlinearities [16], [17], approach integrates DDF and ES-AGSTA, providing rapid,
[18]; however, its inherent chattering effect can be detrimental accurate, and stable responses, as well as the ability to address
to the actuator. Additionally, NMPC was utilized for flow uncertainties.
rate regulation and owned the capability to handle nonlinear The main contributions are highlighted below: (1) A DDF
dynamics and control input constraints [19], [20]; nevertheless, is devised to rapidly drive [Wca , Pca ] close to [Wca,d , Pca,d ],
its heavy computational load poses challenges in ensuring the without requiring model knowledge of the air supply system.
real-time performance of the control system. While these stud- (2) Based on the DDF, a FDC is formulated for [Wca,d , Pca,d ],
ies contribute significantly to the flow rate control, it should which prevents the air supply system from operating in
be noted that they do not consider cathode pressure as a factor the surge zone. (3) The developed ES-AGSTA successfully
in the analysis. eliminates the residual tracking errors of the DDF, with the
Currently, there are several reports on the decoupling control capability to adaptively tune multiple control parameters. (4)
of air supply systems, which can mainly be categorized into The proposed scheme has been implemented in low-cost ECUs

Authorized licensed use limited to: GUANGZHOU UNIVERSITY. Downloaded on February 02,2025 at 10:16:59 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

CHEN et al.: DECOUPLING CONTROL OF FUEL CELL AIR SUPPLY SYSTEM BASED ON DDF AND AGSTA 3

with limited computing power, thus indicating its potential for

industrial applications.
Our data-driven decoupling control methodology is distin-
guished from existing techniques by the complete indepen-
dence from model knowledge, a feature that holds particular
significance for practical applications. The mechanistic model
of an air supply system is notably complex, characterized by
a multitude of parameters, several of which are difficult to
acquire. Furthermore, certain parameters necessitate measure-
Fig. 1. Configuration of air supply system.
ment through sensors, which not only incurs additional costs
but also constrains control performance due to the limitations n ce M O2
in measurement accuracy. In contrast, our DDF-based scheme W O2 ,r e = Ist . (1)
4F
relies solely on the collected data, significantly lowering the The mechanistic model of Wca (Pca , N ) is difficult to acquire.
application threshold. Moreover, even if a flow rate or pressure In a previous study [11], we derived Wca (Pca , N ) through
sensor fails, the feedback control loops can be cut off since polynomial fitting based on a large dataset. However, practical
the DDF can drive the system outputs close to the references. challenges arise in obtaining certain model parameters and
The remainder of the paper is organized as follows. selecting the appropriate order for polynomial fitting. This is
In Section II, the air supply system dynamics and the chal- particularly important due to the varying degrees of nonlinear-
lenges of decoupling control are introduced. Subsequently, the
ity among different types of compressors [27].
control scheme is developed in Section III. Next, Section IV
Remark 2: In [11], we integrated data-driven and model-
discusses the experimental results. Eventually, Section V con-
based approaches, resulting in performance that is reliant
cludes this article.
on the model information. Furthermore, it was influenced
Notation: For the time domain signal u(t) and the transfer
by operating conditions, particularly temperature. Therefore,
function G(s), the following notation is adopted: u(t){G(s)} =
we are interested in exploring alternative methods to effec-
u(t)∗L−1 [G(s)], where L−1 (·) represents the inverse Laplace
tively utilize the dataset, thereby reducing the controller’s
transform, and ’∗’ indicates convolution operator.
reliance on model knowledge. Additionally, our other aim is
to enhance robustness to operating conditions.
II. A IR S UPPLY S YSTEM It is challenging to deduce the explicit relationship between
the outputs and inputs of air supply system from (1). Based
This section discusses the dynamics of the air supply system
on engineering experience, an increase in N leads to a simul-
and the challenges encountered in decoupling control. The air
taneous increase in Wca and Pca . Additionally, an increase in
supply system ensures that the cathode receives an adequate
θ causes an increase in Wca and a decrease in Pca .
supply of oxygen at the appropriate pressure, as depicted in
In general, air supply control encounters the following
Fig. 1. Filtered air is passed through an air compressor to
primary challenges: (1) The strong coupling between Wca and
create a high-pressure air stream, which then passes through
Pca , as well as the inherent nonlinearity of the system, pose
an intercooler, humidifier, and air supply manifold to reach the
cathode. Subsequently, the water produced by the electrochem- a nuisance to decoupling control [7], [28]. (2) There is a
ical reaction is carried out from the cathode outlet by the air lack of effective methods to prevent the air supply system
stream and recycled into a humidifier. Finally, the remaining from entering the surge zone of the compressor. (3) The
exhaust gases are released into the atmosphere through the computational load of the control method needs to be light
throttle. in order to facilitate its implementation on ECUs with limited
Assumption 1: Air compressor speed and throttle angle are computing power.
well regulated and they respond accurately and rapidly to
control commands. III. P ROPOSED C ONTROL S CHEME
Remark 1: Our previous research [26] has developed a This section presents a decoupling scheme for the air supply
throttle control scheme and the air compressor is equipped system, as depicted in Fig. 2, which is mainly composed of a
with a high-precision controller, so Assumption 1 is considered FDC, a DDF, and an ES-AGSTA.
valid.
With Assumption 1, the air supply system can be described A. Data-Driven Feedforward
as [11]:
This part constructs a DDF and presents a FDC. Com-
RTst  pared to physical model-based approaches, DDC demonstrates
Wca (Pca , N ) − Wca,o (Pca , θ )

Ṗca =
Vca Mair greater applicability and flexibility [29], [30]. The complicated
RTst dynamics of the air supply system pose great difficulties for
− W O2 ,r e accurate mathematical modeling [23], thus lookup tables are
Vca M O2

πθ C D,tr A T,tr Pca
 utilized to construct the DDF control:
Wca,o (Pca , θ ) = sin2 √
[N f , θ f ] = f 1 (Wca,d , Pca,d ), f 2 (Wca,d , Pca,d ) . (2)
 
180 RTst
  γ +1
1 2 2(γ −1) The process of data acquisition is outlined as follows. Depend-
· (γ ) 2
γ +1 ing on the characteristics of the test bench, θ and N were

Authorized licensed use limited to: GUANGZHOU UNIVERSITY. Downloaded on February 02,2025 at 10:16:59 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS

Algorithm 1 Feasible Domain Constraint

1: Initialize the parameters [W1 , P1 ], [W2 , P2 ], [W3 , P3 ], [W4 , P4 ]
2: Input [Wca,d , Pca,d ]
3: if Wca,d < min(W1 , W2 , W3 , W4 )
4: Update Wca,d = min(W1 , W2 , W3 , W4 )
5: elseif Wca,d > max(W1 , W2 , W3 , W4 )
6: Update Wca,d = max(W1 , W2 , W3 , W4 )
7: end if
Fig. 2. Overview of the proposed control scheme. 8: if Pca,d < min(P1 , P2 , P3 , P4 )
9: Update Wca,d = min(P1 , P2 , P3 , P4 )
10: elseif Wca,d > max(P1 , P2 , P3 , P4 )
11: Update Wca,d = max(P1 , P2 , P3 , P4 )
12: end if
13: if Pca,d < f (W2 , P2 , W4 , P4 , Wca,d )
14: if Pca,d < f (W2 , P2 , W3 , P3 , Wca,d )
15: Update Pca,d = f (W2 , P2 , W3 , P3 , Wca,d )
16: elseif Pca,d > f (W3 , P3 , W4 , P4 , Wca,d )
17: Update Pca,d = f (W3 , P3 , W4 , P4 , Wca,d )
18: end if
19: elseif Pca,d > f (W2 , P2 , W4 , P4 , Wca,d )
20: if Pca,d < f (W1 , P1 , W2 , P2 , Wca,d )
21: Update Pca,d = f (W1 , P1 , W2 , P2 , Wca,d )
22: elseif Pca,d > f (W4 , P4 , W1 , P1 , Wca,d )
23: Update Pca,d = f (W4 , P4 , W1 , P1 , Wca,d )
24: end if
25: end if

Fig. 3. Lookup tables for DDF: (a) N f = f 1 (Wca,d , Pca,d ); (b)

θ f = f 2 (Wca,d , Pca,d ).

Fig. 4. Schematic of the FDC algorithm.

initially set to 50 deg and 30 krpm (thousand revolutions per
minute), respectively. Subsequently, θ was gradually decreased
in minor increments (2 deg) and the steady-state values of which is achieved by lines 3-12 of Algorithm 1. Subsequently,
[Wca , Pca ] were recorded until the system approached the the FDC (quadrilateral surrounded by four solid lines) is
surge zone. Next, θ was readjusted to 50 deg, N was increased divided into two parts by the red dashed line: its right side
by 10 krpm, and the aforementioned steps were repeated [Pca,d < f (W2 , P2 , W4 , P4 , Wca,d )] and its left side [Pca,d >
until N reaches 90 krpm. The collected data were linearly f (W2 , P2 , W4 , P4 , Wca,d )], which are implemented by lines
interpolated and fitted, resulting in the lookup tables depicted 13-18 and 19-25 of Algorithm 1, respectively.
in Fig. 3. Note that the air supply system exhibits complex non-
In general, a low Wca and high Pca typically lead to linearities, and the DDF is designed through linear fitting,
compressor surges [31]. To address this issue, we pro- which inevitably results in steady-state errors. Additionally,
pose the consideration of constraining their reference DDF alone cannot adapt to varying environmental condi-
trajectories [Wca,d , Pca,d ]. Based on Fig. 3(a), a feasi- tions. Consequently, feedback control is essential to eliminate
ble domain (area enclosed by the solid red line) is steady-state errors and enhance robustness.
derived, where [W1 , P1 ] = [9, 112], [W2 , P2 ] = [25, 107], In contrast to traditional model-based approaches (e.g.,
[W3 , P3 ] = [81, 118], [W4 , P4 ] = [39, 181]. Accordingly, [11]), our method offers advantages in terms of reduced
an FDC is proposed as outlined in Algorithm 1, where dependence on model accuracy and adaptability to changing
f (x A , y A , x B , y B , x) = y A + xy AA −y
−x B (x − x A ), (x A ̸ = x B ).
B operational conditions. As listed in TABLE I, the model-based
Fig. 4 is employed to explain the derivation of the FDC approaches necessitate a variety of model parameters and
algorithm. First, [Wca,d , Pca,d ] is constrained to be between additional temperature sensors, which are not required in our
the minimum and maximum values: method. Additionally, the performance of the model-based
approach degrades significantly when sensors measure delays
or failures in changing operational conditions. On the contrary,
min(W1 , W2 , W3 , W4 ) ≤ Wca,d ≤ max(W1 , W2 , W3 , W4 )
our method possesses enhanced adaptability due to fewer
min(P1 , P2 , P3 , P4 ) ≤ Pca,d ≤ max(P1 , P2 , P3 , P4 ), (3) sensors being employed.

Authorized licensed use limited to: GUANGZHOU UNIVERSITY. Downloaded on February 02,2025 at 10:16:59 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

CHEN et al.: DECOUPLING CONTROL OF FUEL CELL AIR SUPPLY SYSTEM BASED ON DDF AND AGSTA 5

TABLE I
N UMBER OF PARAMETERS AND S ENSORS R EQUIRED

B. ES-AGSTA
This subsection first constructs the GSTA-based feedback
control law and subsequently develops the ES-based adaptive
law.
1) GSTA-Based Control Law: This part constructs the feed-
back control law utilizing GSTA. As mentioned in Section II,
Wca and Pca will rise simultaneously with the compressor
speed increasing, but increasing the throttle angle results in
Fig. 5. Schematic diagram of ES-AGSTA, where ω p ̸= ωq for any
an increase in Wca and a decrease in Pca . Defining the p, q = 1, 2, . . . , 6.
state variables [x1 , x2 ] = [Wca , Pca ], the reference signals
[x1d , x2d ] = [Wca,d , Pca,d ], the tracking errors [e1 , e2 ] = where α ∈ (1/2, 1), and λ1 , λ2 are positive control gains.
[x1 − x1d , x2 − x2d ], and the control inputs [u 1 , u 2 ] = [Nb , θb ], The feedback control signals are constructed utilizing the
then the system dynamics with DDF can be approximated as: GSTA as
ė1 = γ1 u 1 + γ2 θb + D1
 Z t 
1 λ3 2λ3 −1
u1 = −λ1 |s1 | sign(s1 ) − λ2 |s1 | sign(s1 )dτ
ė2 = −γ4 u 2 + γ3 Nb + D2 , (4) γ1
 Z0 t 
1
where γi > 0 for any i = 1, 2, 3, 4, and D1 , D2 denote the u2 = λ6
−λ4 |s2 | sign(s2 ) − λ5 |s2 |2λ6 −1
sign(s2 )dτ .
model errors. γ4 0
Remark 3: Several studies [23], [24], [25] have indicated (10)
that the flow rate is more influenced by compressor speed,
According to Lemma 1, the tracking errors [e1 , e2 ] will
while pressure is more affected by throttle angle. As a result,
converge to a neighborhood to the origin. Defining βi =
Nb and θb are utilized to regulate e1 and e2 , respectively.
λi /γ1 (i = 1, 2), βi = λi /γ4 (i = 4, 5), and βi = λi (i =
It is rewritten as
3, 6), then (10) is rewritten as
ė1 = γ1 u 1 + d1 Z t
ė2 = −γ4 u 2 + d2 , (5) u 1 = −β1 |s1 |β3 sign(s) − β2 |s1 |2β3 −1 sign(s1 )dτ
0
Z t
where d1 = γ2 θb + D1 , d2 = γ3 Nb + D2 are considered as β6
u 2 = −β4 |s2 | sign(s) − β5 |s2 |2β6 −1 sign(s2 )dτ, (11)
bounded lumped disturbances. Additionally, the time deriva- 0
tives of the lumped disturbances [ḋ1 , ḋ2 ] are bounded from an where β1 , β2 , β4 , β5 > 0, and 1/2 < β3 , β6 < 1. The control
energy perspective. law (10) contains unknown parameters (γ1 , γ4 ), which are
Since the relative degrees of [e1 , e2 ] with regard to [u 1 , u 2 ] therefore incorporated into βi and subsequently tuned by ES.
are one, the sliding variables are constructed as Additionally, considering the presence of measurement
[s1 , s2 ] = [e1 , −e2 ], (6) noise, an ET integrator is designed as
Z t
which yields  x(τ )dτ, if |e(t)| > ET
y(t) = (12)
ṡ1 = γ1 u 1 + d1  0
y(t− ), if 0 ≤ |e(t)| ≤ ET,
ṡ2 = γ4 u 2 − d2 . (7)
where y(t− ) represents the integrator state at the previous
Lemma 1: [32] For a system in the presence of a distur- control period, and e(t) denotes the current tracking error.
bance d(t): 2) ES-Based Adaptive Law: This part develops an ES-based
adaptive law and demonstrates its gradient descent principle
ṡ1 = h(t)u + d(t), (|ḋ(t)| ≤ K d ), (8) and stability. In reality, there is a game playing between |e1 |
the following GSTA ensures that the sliding variable (s1 ) and |e2 |, which poses a challenge for manual parameter tuning.
converges in finite time: Accordingly, ES was considered for the development of an
 ES-GASTA (Fig. 5) due to its model-free advantage.
 u = 1 −λ |s |α sign(s ) + v  Remark 4: Another benefit of the ET integrator (12) is that
1 1 1
h(t) (9) it will cause the meaningless game to cease when the tracking
v̇ = −λ2 |s1 |2α−1
sign(s1 ),

error falls within the specified ET.

Authorized licensed use limited to: GUANGZHOU UNIVERSITY. Downloaded on February 02,2025 at 10:16:59 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS

(1) The perturbation amplitudes (ai > 0) should be tiny

to ensure accurate estimation, which are typically two
orders of magnitude smaller than the target range of the
optimized variable.
(2) The perturbation frequencies (ωi ) must ensure a time-scale
separation between the ES loops and the system to be
optimized: max(ωi ) ≤ 5T 2π
. Moreover, to preclude bias
terms arising from demodulation in the case of a quadratic
nonlinearity, another principle is written as
ω p + ωq ̸ = ωr , for any p, q, r = 1, 2, . . . , 6. (17)
(3) The cut-off frequency (h) is lower than the minimum
Fig. 6. Schematic representation of ES procedure. perturbation frequency: 0 < h < min(ωi ).
(4) The update gains (µi > 0) are gradually increased from
Remark 5: Adaptive control is capable of effectively man- a small value until a satisfactory convergence rate is
aging uncertainty, enhancing robustness, flexibility, and control achieved. Additionally, the update gains remain at zero
accuracy, as well as reducing parameter tuning time [33], until the ES loops are activated.
[34], [35]. Mei et al. [32] only provide an empirically based Next we discuss the rationale behind the selection guidelines
parameter selection scheme for GSTA, so the ES-AGSTA point by point:
serves as a valuable complement to their work. (1) According to Theorem 2, ai are directly related to the
For writing conciseness, the following notations are utilized: size of the set to which the ES will converge. Additionally,
 it is too small leading to insufficient excitation and too
B = [β1 , β2 , · · · , β6 ]T
large inducing notable variations in the control parameters.


 T
B = β1 , β2 , · · · , β6∗
 ∗  ∗ ∗
Consequently, it is advisable to set ai two orders of magnitude



 h iT
 B̂ = β̂1 , β̂2 , · · · , β̂6

 smaller than the latter’s target range.
h iT (13) (2) ES optimizes the control performance based on the



 B̃ = B ∗
− B̂ = β̃ 1 , β̃2 , · · · , β̃6 cost function (14), thus T determines the time scale of the

 A = [A1 , A2 , · · · , A6 ]T

 optimized system. Therefore, the perturbation periods must be
 notably longer than T , which ensures that B is considered
= [a1 sin ω1 t, a2 sin ω2 t, · · · , a6 sin ω6 t]T .


constant over an iteration so that J (m, B) approaches the
To evaluate the control performance, a cost function is formu- steady-state cost. Normally, it is suggested that the pertur-
lated as bation periods be no less than 5T [39]: 2π /ωi ≥ 5T (i =
Z mT    1, 2, · · · , 6), which leads to max(ωi ) ≤ 5T 2π
. Furthermore,
q 0 e1
J (m, B) = [e1 , e2 ] 1 dt (m ≥ 1). (14) i (t) described in (33) can not contain constant terms to
(m−1)T 0 q 2 e2
ensure the stability of the estimation error system (38) at
Assumption 2: The cost function J (B) has a minimum J ∗ B ∗ . Accordingly, the perturbation frequencies need to satisfy
at B = B ∗ . ω p + ωq ̸ = ωr , for any p, q, r = 1, 2, . . . , 6.
Remark 6: J (B) denotes the accumulation of tracking (3) Hh (s) = s+h s
is utilized to remove the DC component
errors in one iteration period, so it possesses at least one while preserving all perturbations [36]. Hence, h is lower than
minimum. the minimum frequency: 0 < h < min(ωi ).
ES employs sinusoidal perturbation signals (a sin ωi t) to (4) The convergence rate of β̂i is positively relative to µi
extract the local gradient, and then utilizes a gradient descent as shown in (24), but overly large µi may lead to system
method to reduce J (B), as depicted in Fig. 6. instability [37]. As a result, µi shall start from a small positive
The equations governing the i-th (i = 1, 2, . . . , 6) ES loop number and gradually increase to accelerate the convergence
in Fig. 5 are as follows: rate while maintaining ES stability. Additionally, µi should
remain at zero before the ES loops are activated to prevent β̂i
βi = β̂i + ai sin ωi t from deviating from their initial values.
n µ o
i
β̂i = ξi − Remark 8: The perturbation frequencies affect the dynam-
s ics and performance of the overall system, primarily in the
ξi = η sin ωi t following aspects: (1) The convergence rate of the ES relates
 
s positively to the frequency. The gradient information can be
η = J (B) . (15)
s+h extracted faster as the frequency increases, thus accelerating
the seeking process. (2) Conversely, the stability of system
where Hh (s) = s+h
s
is a washout filter and − µsi is an integrator response is negatively correlated with frequency. An increase
with update gain. The estimation errors are defined as in frequency results in more rapid variations in control param-
β̃i = βi∗ − β̂i (i = 1, 2, . . . , 6). (16) eters, which leads to frequent fluctuations in control input. (3)
Computational load is positively associated with frequency,
Remark 7: The hyperparameters are particularly critical for since higher frequency may require more frequent sampling
ES, whose selection guidelines are summarized below [36], and computation. (4) If any perturbation frequency approaches
[37], [38]: the system’s natural frequencies, resonance phenomena may
Authorized licensed use limited to: GUANGZHOU UNIVERSITY. Downloaded on February 02,2025 at 10:16:59 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

CHEN et al.: DECOUPLING CONTROL OF FUEL CELL AIR SUPPLY SYSTEM BASED ON DDF AND AGSTA 7

be induced, leading to significant oscillations or instability Similarly, for the ES loop2 to ES loop6, one can obtain
within the system. Therefore, it is essential to ensure that µi ai ∂ J
the perturbation frequencies are sufficiently distanced from β̂˙i = − · , (i = 2, 3, · · · , 6). (24)
the natural frequencies. Fortunately, our approach satisfies this 2 ∂βi βi =β̂i
requirement in principle. Consequently, B̂ will converge to B ∗ in a gradient descent
Theorem 1: For the ES loops in Fig. 5, if the guidelines manner since µi ai > 0.
listed in Remark 7 are followed, then B̂ will converge to B ∗ The proof is complete.
in a gradient descent manner. Remark 9: The significant advantage of ES is the low
Proof: Taylor expansion is performed on J (B) at B̂ computational burden, making it possible for implementation
and second-order and above terms are neglected. Here B̂ on ECUs with limited computing power. However, ES enforces
is not considered to be equal to B ∗ , thus the first-order the gradient to be equal to zero, which means that it can
terms dominate compared to the higher-order ones. Combined merely secure a local minimum instead of the global one.
with (15), one gets A common remedy is to run the ES multiple times with
β1 − β̂1 different initial conditions [37]. Furthermore, the local optimal
 
β2 − β̂2  solution typically yields satisfactory control outcomes without
 · · ·  = J ( B̂)
J (B) ≈ J ( B̂) + ∇ J ( B̂)  
necessitating pursuit of the global one.
β6 − β̂6 Theorem 2: For the system in Fig. 5, following the guide-

a1 sin ω1 t
 P error B̃ achieves local
lines listed in Remark 7, the estimated
exponential convergence to an O(6 6p=1 a 2p ) neighborhood of
a2 sin ω2 t 
+ ∇ J ( B̂) 
 ··· ,
 (18) the origin.
Proof: From (16) it derives that
a6 sin ω6 t
β̂i = βi∗ − β̃i , (25)
where the gradient is denoted as
" # which is plugged into βi = β̂i + ai sin ωi t yields
∂J ∂J ∂J
∇ J ( B̂) = , ,··· , .
∂β1 β1 =β̂1 ∂β2 β2 =β̂2 ∂β6 β6 =β̂6 βi − βi∗ = ai sin ωi t − β̃i , (26)
(19) in other words
Accordingly, the perturbation signals (ai sin ωi t) are modu- B − B ∗ = A − B̃. (27)
lated by the local gradient ∇ J ( B̂), as shown in Fig. 6.
Coherent demodulation is utilized for extracting the gradi- Taylor expansion is performed on J (B) at the minimum point
ent. Let us commence the analysis with ES loop1. Hh (s) has B ∗ and higher-order terms are neglected:
zero DC gain, which removes the DC component J ( B̂): J (B) ≈ J (B ∗ ) + (B − B ∗ )T R(B − B ∗ )
a1 sin ω1 t
 
= J ∗ + ( B̃ − A)T R( B̃ − A), (28)
a2 sin ω2 t 
 ··· .
η = J (B) {Hh (s)} = ∇ J ( B̂)   (20) where R6×6 = (R pq )6×6 ( p, q = 1, 2, . . . , 6) is symmetric
a6 sin ω6 t positive definite since it is equal to the Hessian matrix divided
by 2. The gradient at B ∗ is equal to zero, so the second-order
Then applying sin α · sin β = 1
2 [cos(α − β) − cos(α + β)] Taylor expansion is utilized.
yields Replacing β̂i from (15), (16) is rewritten as
nµ o
a1 sin2 ω1 t
 
i
a2 sin ω2 t · sin ω1 t  β̃i = βi∗ + sin ωi t {Hh (s)} J (B). (29)
s
ξ1 = η sin ω1 t = ∇ J ( B̂)  
 ···  Plugging (28) into (29) yields
a6 sin ω6 t · sin ω1 t nµ o
i
 a1
(1 − cos 2ω1 t)
 β̃i = βi∗ + sin ωi t {Hh (s)} [J ∗ + ( B̃ − A)T R( B̃ − A)].
2 s
 a2 [cos(ω1 − ω2 )t − cos(ω1 + ω2 )t] (30)
= ∇ J ( B̂)  2  (21)
 ··· 
a6 Expanding ( B̃ − A)T R( B̃ − A) in (30) produces
2 [cos(ω1 − ω6 )t − cos(ω1 + ω6 )t]
6 X
6
∂J a1
Since ξ1 contains the DC component ( ∂β · 2 ), the
X
1 β1 =β̂1 ( B̃− A)T R( B̃− A) = R pq (β̃ p β̃q + A p Aq − β̃ p Aq − β̃q A p ).
sinusoidal components are filtered out due to the infinite DC p=1 q=1
gain of the integrator (− µs1 ): (31)
n µ o ∂J a1 n µ1 o
1
β̂1 = ξ1 − ≈ · − . (22) Lemma 2: [40] If the transfer function H (s) has all of its
s ∂β1 β1 =β̂1 2 s poles with negative real parts, then
Rearranging (22) and calculating the time derivative produces h i
{H (s)} sin ωt = Im H ( jω)e jωt + ϵ −t ,
µ1 a1 ∂ J
β̂˙1 = − · . (23)
2 ∂β1 β1 =β̂1 where ϵ −t denotes exponentially decaying terms.

Authorized licensed use limited to: GUANGZHOU UNIVERSITY. Downloaded on February 02,2025 at 10:16:59 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS

The terms containing J ∗ and A p Aq in (30) are simplified Differentiating (37) with respect to time and noticing that βi∗
utilizing Lemma 2: is constant, one gets
   
6 X
6 6
β̃˙i = µi −ai
X
sin ωi t {Hh (s)}  J ∗ + R pq A p Aq  = i (t) + ϵ −t ,
X
Riq β̃q + i (t) + ϵ −t  . (38)
p=1 q=1 q=1
(32)
The homogeneous part of (38) is considered:
where 6
β̃˙i = −µi ai
X
6 X
X 6 Riq β̃q , (i = 1, 2, · · · , 6), (39)
i (t) = R pq a p aq {H pq1 sin[(ωi + ω p + ωq )t − M pq1 ] q=1
p=1 q=1
whose state-space representation is written as
+ H pq2 sin[(ωi + ω p − ωq )t − M pq2 ]
+ H pq3 sin[(ωi − ω p + ωq )t − M pq3 ] B̃˙ = C B̃, (40)
+ H pq4 sin[(ωi − ω p − ωq )t − M pq4 ]}, (33)
where
where
µ1 a 1
  
( O R11 R12 ··· R16
H pq1 = H pq4 = |Hh [ j (ω p + ωq )]|  µ2 a 2   R21 R22 ··· R26 
(34) C = −

..   ..

.. .. .. 

(41)
H pq2 = H pq3 = |Hh [ j (ω p − ωq )]|,  .  . . . . 
and the constants M pqr (r = 1, 2, 3, 4) are determined by the O µ6 a 6 R61 R62 · · · R66
phase of Hh [ j (ω p ± ωq )]. Since µi ai > 0 and R is symmetric positive definite,
Remark 10: i (t) does not contain P constant
P6terms accord- it can be determined that C is a Hurwitz-stable matrix
ing to (17), and is of order O( 6p=1 q=1 a p aq ) = according to Sylvester’s criterion [41], which indicates expo-
P6
O(6 p=1 a p ).2
nential stability of (40). Furthermore, as the persistent part
Combining (31) and (32), exploiting symmetry of R and of the
P non-homogeneous forcing term P in (38) is of order
noticing that h < ω p , we rewrite (30) after dropping out O(6 6p=1 a 2p ), B̃ converges to an O(6 6p=1 a 2p ) neighbor-
the second-order terms containing β̃ p β̃q since a local result hood of the origin.
is being proved: Remark 11: The exponential stability of the homogeneous
nµ o
i
part of (38) implies L-stability of the system with forcing.
β̃i = βi∗ + Furthermore, the bounded inhomogeneous forcing leads to the
s
 
6 X 6
  boundedness of the solution of the forced system [42].
sin ωi t 
X
−2R pq β̃q A p  + i (t) + ϵ −t  . The proof is complete.
Remark 12: Briefly, ES-GSTA first defines a cost function
p=1 q=1
J based on the tracking error, and subsequently adapts to
(35) disturbances and uncertainties in real-time by reducing J .
Noting that A p = a p sin ω p t, one gets In this process, the control parameters are updated to seek the
  minimum value of J . Different disturbances and uncertainties
nµ o 6 X6 commonly correspond to different minimum points, which in
i
X
sin ωi t  −2R pq β̃q A p  turn lead to different control parameters. Unlike model-based
s adaptive methods (e.g., some observers), it does not rely on
p=1 q=1
nµ 6 X
oX 6 any knowledge about the plant. Our future work will aim to
=−
i
2R pq β̃q a p sin ω p t · sin ωi t extend it to general nonlinear systems.
s
p=1 q=1
6 X
nµ o X 6 IV. R ESULTS AND D ISCUSSION
i
=− R pq β̃q a p [cos(ω p − ωi )t − cos(ω p + ωi )t]
s In this section, the proposed method is sequentially vali-
p=1 q=1
dated on a test bench for air supply system and a HiL platform
nµ o 6
i
X for FCEV.
=− ai Riq β̃q , (36)
s
q=1

since only the terms containing cos(ω p − ωi )t = 1 ( p = A. Test Bench Experimental Results
i) are preserved, as the integrator ( µsi ) has infinite DC gain. This subsection validates the proposed method on an air
Plugging (36) into (35) results in supply system test bench, as depicted in Fig. 7, which is
  equipped with a RapidECU-F2 (200 MHz dual-core micro-
nµ o 6
i 
X processor) from Huahai Technologies. The proposed scheme
β̃i = βi∗ + −ai Riq β̃q + i (t) + ϵ −t  . (37) is developed in MATLAB/Simulink, which is built in a PC
s
q=1 and then downloaded to the ECU via a CAN bus.

Authorized licensed use limited to: GUANGZHOU UNIVERSITY. Downloaded on February 02,2025 at 10:16:59 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

CHEN et al.: DECOUPLING CONTROL OF FUEL CELL AIR SUPPLY SYSTEM BASED ON DDF AND AGSTA 9

Fig. 7. Test bench for air supply system.

Fig. 8. Tracking responses and cost function during parameter optimization.
TABLE II
PARAMETERS FOR T EST B ENCH E XPERIMENTS

1) Parameter Optimization: This part exhibits that the

ES-AGSTA adaptively optimizes the control performance. The
experimental parameters were presented in Table II. The ETs
are determined by the measurement noises of [Wca , Pca ]. Fig. 9. Iterative update procedure for the controller parameters.
In addition, depending on the dynamics of the air supply
system, T is chosen to be 10 seconds, which also serves
as the period of the reference signals used for optimization. J and B̂ converged at 3.34 and [0.148, 0.076, 0.705, 0.179,
Furthermore, q1 = q2 means that the tracking accuracy of Wca 0.036, 0.714], respectively.
and Pca is considered with equal importance. Moreover, the 2) Control Performance Evaluation: This part evaluates
ES hyperparameter settings follow the guidelines in Remark 7. the performance of the proposed scheme. The optimized
Specific details are listed below: parameters were loaded into the controller and the ES loops
(1) ai shall be two orders of magnitude smaller than the were disabled to reduce the computational load. To adequately
target range of the control parameters. They were selected as evaluate the decoupling capability, two sets of simultaneously
1 × 10−3 since the latter are of the order of 10−1 . varying reference trajectories were selected. Case 1: Wca,d
(2) ωi first follow max(ωi ) ≤ 5T 2π 2π
= 5×10 ≈ 0.1257 rad/s, and Pca,d were chosen as a step signal and a sinusoidal signal,
so we let max(ωi ) = 0.093 rad/s. Second, they also respectively, with the results depicted in Fig. 10. Case 2: Wca,d
necessitate to satisfy ω p + ωq ̸ = ωr for any p, q, r = and Pca,d were selected as a sinusoidal signal and a step signal,
1, 2, . . . , 6. Accordingly, [ω1 , ω2 , · · · , ω6 ] were finally picked respectively, with the results depicted in Fig. 11.
as [9.3,8.3,7.3,6.3,5.3,4.3]×10−2 rad/s. According to the experimental results, the proposed method
(3) h needs to be smaller than the minimum perturbation effectively completes the decoupling of Wca and Pca :
frequency: 0 < h < 4.3 × 10−2 rad/s, and retain a certain (1) Even if [Wca,d , Pca,d ] varied simultaneously, they were
interval. Therefore, h was elected to be 1 ×10−2 rad/s. still tracked quickly and accurately by [Wca , Pca ], as depicted
(4) µi are gradually increased from a small value until a in Figs. 10(a,f) and 11(a,f). The step responses for both
satisfactory convergence rate is achieved. They were selected Wca and Pca are completed within 0.5 s [Figs. 10(a)
as 1 × 10−2 after a couple of attempts. and 11(f)]. When tracking sinusoidal signals, the absolute
(5) The initial values [β̂1 (0), β̂2 (0), · · · , β̂6 (0)] were set as errors [|e1 |, |e2 |] are limited to about [2 g/s, 1 kPa], as depicted
[0.2,0.1,0.7,0.2,0.1,0.7] after several trails based on engineer- in Figs. 11(b) and 10(g).
ing experience. (2) The jitter of Pca is weak when Wca undergoes a step
The ES-AGSTA reduced J (m, B) online for more favorable change [Fig. 10(a,f), 102-103 s]. However, there is signif-
tracking responses [Fig. 8], which resulted in the iterative icant jitter in Wca when the Pca undergoes a step change
update of the control parameters depicted in Fig. 9. Eventually, [Fig. 11(a,f), 313-314 s]. This phenomenon is attributed to the

Authorized licensed use limited to: GUANGZHOU UNIVERSITY. Downloaded on February 02,2025 at 10:16:59 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS

Fig. 10. Control performance in case 1: (a) flow rate response; (b) flow rate error; (c) N and Nact ; (d) N f ; (e) Nb ; (f) pressure response; (g) pressure
error; (h) θ and θact ; (i) θ f ; (j) θb .

Fig. 11. Control performance in case 2: (a) flow rate response; (b) flow rate error; (c) N and Nact ; (d) N f ; (e) Nb ; (f) pressure response; (g) pressure
error; (h) θ and θact ; (i) θ f ; (j) θb .

TABLE III
RMSE S FOR C ONTROL T ECHNOLOGIES

Fig. 12. Comparison with the DMD, ODMD, and ASTA on the test bench:
(a) tracking responses in case 1; (b) tracking responses in case 2.

characteristics of the air supply system, as the relative degree

of Pca with regard to Wca is one [refer to (1)].
(3) In the proposed scheme, [N f , θ f ] dominate the variation The quantitative analysis was carried out utilizing root mean
trend of [N , θ] [Fig. 10(d,i)], while [Nb , θb ] play a role in fine- square error (RMSE):
tuning [Fig. 10(e,j)]. Specifically, [N f , θ f ] drive [Wca , Pca ] to s
the vicinity of [Wca,d , Pca,d ], and [Nb , θb ] is responsible for
PN  2
Wca (i) − Wca,d (i)
i=1
removing the residual tracking errors. RMSE1 =
N
To further declare the superior performance of our approach, s
PN  2
we compared it with the following techniques: a DMD [23], i=1 Pca (i) − Pca,d (i)
an optimization DMD (ODMD) [24]. Additionally, an adaptive RMSE2 = , (42)
N
supertwisting algorithm (ASTA) [43] was employed to replace
the ES-AGSTA as an additional comparison. The results are where N is the number of datasets, with the results listed in
depicted in Fig. 12. Table III.

Authorized licensed use limited to: GUANGZHOU UNIVERSITY. Downloaded on February 02,2025 at 10:16:59 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

CHEN et al.: DECOUPLING CONTROL OF FUEL CELL AIR SUPPLY SYSTEM BASED ON DDF AND AGSTA 11

TABLE V
RMSE S FOR C ONTROL T ECHNOLOGIES

TABLE VI
C OMPUTATIONAL C OSTS OF THE M ETHODS

Fig. 13. HiL platform for FCEV.

TABLE IV
PARAMETERS FOR H I L E XPERIMENTS

values [Fig. 14(c,d)]. To demonstrate the advantages of our

approach, the following techniques were used for compar-
ison: a DDC [25], an ODMD [24], and an ASTA [43].
It is noteworthy that FDC is also integrated into the three
techniques. Their RMSEs are listed in Table V, indicating that
our method tracks the reference trajectories more accurately.
Furthermore, the proposed scheme demonstrates reduced over-
shooting compared to ODMD, with almost the same response
In summary, the proposed method achieves decoupling con- time. [Fig. 15(a,c)]. Moreover, it exhibits a faster response
trol of Wca and Pca and offers advantages in tracking accuracy speed than DDC and ASTA [Fig. 15(a,c)].
compared with the DMD, ODMD, and ASTA techniques. In the overall control strategy integration, the FDC should
be positioned after the calculated [Wca,d , Pca,d ] (43), and
before the DDF as well as ES-AGSTA (see Fig. 2). In addi-
B. HiL Experimental Results tion, the FDC in our method was removed as an additional
This section further validates the proposed approach on comparison to demonstrate the effectiveness of the FDC in
a FCEV utilizing the HiL platform. It primarily comprises the surge prevention under sudden changes in stack current,
an ECU (200 MHz dual-core microprocessor), a dSPACE with the results depicted in Fig. 16. In the absence of FDC,
(2.8 GHz quad-core processor), and a PC, as illustrated in Pca,d remained at 151.95 kPa (PR was reset to 1.5). As the
Fig. 13. Specifically, the ECU functions as the controller while Wca drops, the system enters the surge zone, causing drastic
the dSPACE carries a high-fidelity commercial FCEV model and irregular fluctuations in flow rate and pressure [Fig. 16(a)].
as the plant. The parameters for the HiL experiments are In the presence of FDC, Pca,d is dynamically adjusted accord-
detailed in Table IV. ing to Wca,d [Fig. 16(b)], which in turn avoids the surge
The controller parameter optimization is similar to that phenomenon.
of the previous subsection, and this part highlights the per- Computational costs deserve attention as well. We ran the
formance of the proposed method during FCEV operation. above controllers for 1 × 106 iterations on a desktop computer
The driving condition was selected to cyclic UDDS (urban (with an Intel Core i9-13900K at 3.0 GHz processor) and
dynamometer driving schedule) [Fig. 14(a)], and a rule-based recorded the computation times as listed in Table VI, since
energy management strategy was developed, which resulted the real-time system ECU could not capture the distinction in
in the stack current as illustrated in Fig. 14(b). The calculated computational cost. Our method possesses remarkably shorter
[Wca,d , Pca,d ] [see Fig. 14(c,d)] were then deduced from the computation time for the following main reasons: (1) It
OER and PR [44]: has significantly simpler composition (only feedforward and
OER n ce M O2 feedback) compared to the DDC. (2) In contrast to the ODMD,
Wca,d = · Ist , Pca,d = PR · Patm . (43) there are no four higher-order decoupled transfer functions. (3)
y O2 4F
In comparison with the ASTA, the optimizer can be disabled
Furthermore, the following varying conditions were considered after the parameter optimization is complete.
to comprehensively assess the robustness: stack temperature, Overall, the proposed method prevents the air supply system
relative humidity, load power [Fig. 14(e,f,g)], as well as from entering the surge zone and demonstrates robustness
fluctuations in anode pressure due to purging [Fig. 14(h)]. to varying operating conditions. Besides, it offers advantages
Subject to the FDC, the actual [Wca,d , Pca,d ] [see in tracking accuracy, response speed, overshoot suppression,
Fig. 15(a,c), cyan line] were derived from the calculated as well as computational cost.

Authorized licensed use limited to: GUANGZHOU UNIVERSITY. Downloaded on February 02,2025 at 10:16:59 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

12 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS

Fig. 14. Operating conditions of the FCEV: (a) vehicle velocity; (b) stack current; (c) calculated Wca,d ; (d) calculated Pca,d ; (e) stack temperature Tst ;
(f) stack relative humidity; (g) stack load power; (h) fluctuations in anode pressure.

Fig. 15. Comparison with the DDC, ODMD and ASTA on the FCEV: (a) flow rate response; (b) flow rate error; (c) pressure response; (d) pressure error.

achieving more favorable control performance. Experiments

have demonstrated the effectiveness and advantages of our
approach.
The significant difference lies in its independence from
model knowledge and its effective avoidance of compressor
surge, as compared with existing reports. The proposed decou-
pling method is easily portable, with potential for industrial
applications. Additionally, we also present a multi-parameter
tuning scheme for the GSTA.
Note that a limitation of this study is the focus only on air
supply control without considering fuel cell efficiency, which
is greatly affected by the air supply system. It is of interest
to optimize [Wca,d , Pca,d ] to maximize the efficiency, which
Fig. 16. Anti-surge performance test of FDC: (a) without FDC; (b) with will also be explored in future research.
FDC.

V. C ONCLUSION A PPENDIX
This article has investigated a model-free decoupling This part discusses considerations for real-world deploy-
method for the air supply system, which achieves the inde- ments of our approach.
pendent regulation of air flow rate and pressure. The main 1. Hardware requirements: In this study, the mass-produced
findings are as follows: (1) The proposed method does not RapidECU-F2 was utilized to run the proposed scheme, with
rely on model knowledge and exhibits fine portability. (2) the main technical specifications listed in Table VII. It
The suggested FDC effectively prevents the air supply system guarantees that our method operates reliably, so we recom-
from entering the surge zone, thus providing protection for mend referring to the specifications for industrial applications.
the compressor. (3) The presented ES-AGSTA possesses the Additionally, a gas flow rate sensor and a gas pressure sensor
capability for multi-parameter tuning, which contributes to are required.

Authorized licensed use limited to: GUANGZHOU UNIVERSITY. Downloaded on February 02,2025 at 10:16:59 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

CHEN et al.: DECOUPLING CONTROL OF FUEL CELL AIR SUPPLY SYSTEM BASED ON DDF AND AGSTA 13

TABLE VII [10] R. J. Talj, D. Hissel, R. Ortega, M. Becherif, and M. Hilairet, “Exper-
T ECHNICAL S PECIFICATIONS FOR R APID ECU-F2 imental validation of a PEM fuel-cell reduced-order model and a
moto-compressor higher order sliding-mode control,” IEEE Trans. Ind.
Electron., vol. 57, no. 6, pp. 1906–1913, Jun. 2010.
[11] M. Li, H. Yin, T. Ding, J. Zhao, J. Gao, and H. Chen, “Air flow rate and
pressure control approach for the air supply subsystems in PEMFCs,”
ISA Trans., vol. 128, pp. 624–634, Sep. 2022.
[12] I. M. M. Saleh, R. Ali, and H. Zhang, “Simplified mathematical model
of proton exchange membrane fuel cell based on horizon fuel cell
stack,” J. Modern Power Syst. Clean Energy, vol. 4, no. 4, pp. 668–679,
Oct. 2016.
[13] J. Chen, Z. Liu, F. Wang, Q. Ouyang, and H. Su, “Optimal oxygen excess
2. Computational demands: Our approach possesses low ratio control for PEM fuel cells,” IEEE Trans. Control Syst. Technol.,
vol. 26, no. 5, pp. 1711–1721, Sep. 2018.
computational demands, mainly from control loops and ES
[14] Y. Ma, F. Zhang, J. Gao, H. Chen, and T. Shen, “Oxygen excess ratio
loops, the latter being more computationally intensive. In our control of PEM fuel cells using observer-based nonlinear triple-step
previous study [26], the RapidECU-F2 successfully ran thir- controller,” Int. J. Hydrogen Energy, vol. 45, no. 54, pp. 29705–29717,
teen ES loops with a sampling period of 5 ms. After testing, Nov. 2020.
[15] Y. Hu, H. Chen, X. Gong, S. Yu, J. Gao, and H. Chen, “Control-
the RapidECU-F2 can run the proposed method at 2 ms oriented modeling and robust nonlinear triple-step controller design for
sampling period. an air-feed system for polymer electrolyte membrane fuel cells,” Asian
3. Safety considerations: (1) The anode pressure controller J. Control, vol. 21, no. 4, pp. 1811–1823, Jul. 2019.
needs to be enabled during data acquisition to ensure proper [16] J. Liu, Y. Gao, X. Su, M. Wack, and L. Wu, “Disturbance-observer-
based control for air management of PEM fuel cell systems via sliding
pressure difference on both sides of the PEM. (2) Considera- mode technique,” IEEE Trans. Control Syst. Technol., vol. 27, no. 3,
tion could be given to adding an additional safety mechanism pp. 1129–1138, May 2019.
that cuts off the feedback control loops when a sensor anomaly [17] A. Pilloni, A. Pisano, and E. Usai, “Observer-based air excess ratio
is detected. DDF can drive [Wca , Pca ] close to [Wca,d , Pca,d ]. control of a PEM fuel cell system via high-order sliding mode,” IEEE
Trans. Ind. Electron., vol. 62, no. 8, pp. 5236–5246, Aug. 2015.
4. Potential challenges for implementation in low-power
[18] W. Garcia-Gabin, F. Dorado, and C. Bordons, “Real-time implementa-
environments: (1) If computational resources are more limited, tion of a sliding mode controller for air supply on a PEM fuel cell,”
it may not be possible to perform ES computations. A feasible J. Process Control, vol. 20, no. 3, pp. 325–336, Mar. 2010.
solution is to separate the ES loops from the control loops [19] Q. Ouyang, J. Chen, F. Wang, and H. Su, “Nonlinear MPC controller
and lengthen the sampling period for the former [39]. (2) design for AIR supply of PEM fuel cell based power systems,” Asian
J. Control, vol. 19, no. 3, pp. 929–940, May 2017.
After the optimization is complete, ES loops still consume [20] J. K. Gruber, C. Bordons, and A. Oliva, “Nonlinear MPC for the airflow
computational resources. Accordingly, it is recommended to in a PEM fuel cell using a Volterra series model,” Control Eng. Pract.,
disable ES loops after the optimized parameters are loaded vol. 20, no. 2, pp. 205–217, Feb. 2012.
into the controller. [21] D. Zhao, L. Xia, H. Dang, Z. Wu, and H. Li, “Design and control of air
supply system for PEMFC UAV based on dynamic decoupling strategy,”
Energy Convers. Manage., vol. 253, Feb. 2022, Art. no. 115159.
R EFERENCES [22] H. Yuan, H. Dai, P. Ming, J. Zhan, X. Wang, and X. Wei, “A fuzzy
extend state observer-based cascade decoupling controller of air supply
for vehicular fuel cell system,” Energy Convers. Manage., vol. 236,
[1] H. Hu, K. Ou, and W.-W. Yuan, “Fused multi-model predictive
May 2021, Art. no. 114080.
control with adaptive compensation for proton exchange mem-
brane fuel cell air supply system,” Energy, vol. 284, Dec. 2023, [23] T. Zeng et al., “Feedforward-based decoupling control of air supply for
Art. no. 128459. vehicular fuel cell system: Methodology and experimental validation,”
[2] D.-M. Kim, J.-W. Chin, J.-H. Kim, and M.-S. Lim, “Analytical temper- Appl. Energy, vol. 335, Apr. 2023, Art. no. 120756.
ature estimation process of the air supply system of the proton exchange [24] Y. Qiu et al., “Coupling mechanism analysis and decoupling control
membrane fuel cell stack in fuel cell electric vehicles,” Energy, vol. 283, of the air supply system for fuel cell engine in fuel cell vehicle,”
Nov. 2023, Art. no. 129053. IEEE Trans. Transport. Electrific., vol. 10, no. 4, pp. 10059–10072,
[3] G. Lu et al., “Study on counter-flow mass transfer characteris- Dec. 2024.
tics and performance optimization of commercial large-scale proton [25] C. Zhang, Y. Hu, X. Gong, Y. Huang, and H. Chen, “Design and
exchange membrane fuel cells,” Appl. Energy, vol. 359, Apr. 2024, experimental verification of model-free adaptive sliding controller for air
Art. no. 122743. supply system of PEMFCs,” Control Eng. Pract., vol. 128, Nov. 2022,
[4] H. Chen, Z. Liu, X. Ye, L. Yi, S. Xu, and T. Zhang, “Air flow and Art. no. 105336.
pressure optimization for air supply in proton exchange membrane fuel [26] L. Chen, Y. Wang, J. Zhao, S. Ding, J. Gao, and H. Chen, “A novel
cell system,” Energy, vol. 238, Jan. 2022, Art. no. 121949. adaptive control scheme for automotive electronic throttle based on
[5] H. Chen, B. Liu, T. Zhang, and P. Pei, “Influencing sensitivities of extremum seeking,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 70,
critical operating parameters on PEMFC output performance and gas no. 6, pp. 2599–2611, Jun. 2023.
distribution quality under different electrical load conditions,” Appl. [27] J. Hou, M. Yang, C. Ke, and J. Zhang, “Control logics and strategies for
Energy, vol. 255, Dec. 2019, Art. no. 113849. air supply in PEM fuel cell engines,” Appl. Energy, vol. 269, Jul. 2020,
[6] Y. Li, X. Zhao, S. Tao, Q. Li, and W. Chen, “Experimental study on Art. no. 115059.
anode and cathode pressure difference control and effects in a proton [28] Z. Liu, B. Zhang, and S. Xu, “Research on air mass flow-pressure com-
exchange membrane fuel cell system,” Energy Technol., vol. 3, no. 9, bined control and dynamic performance of fuel cell system for vehicles
pp. 946–954, Sep. 2015. application,” Appl. Energy, vol. 309, Mar. 2022, Art. no. 118446.
[7] Z. Liu, H. Chen, L. Peng, X. Ye, S. Xu, and T. Zhang, “Feedforward- [29] J. Li, Y. Fang, and L. Zhang, “A TM-based adaptive learning data-
decoupled closed-loop fuzzy proportion-integral-derivative control of air model for trajectory tracking and real-time control of a class of nonlinear
supply system of proton exchange membrane fuel cell,” Energy, vol. 240, systems,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 69, no. 2,
Feb. 2022, Art. no. 122490. pp. 859–871, Feb. 2022.
[8] J. T. Pukrushpan, Modeling and Control of Fuel Cell Systems and Fuel [30] W. Zhang, S. Mao, J. Huang, L. Kocarev, and Y. Tang, “Data-driven
Processors. Ann Arbor, MI, USA: Univ. Michigan, 2003. resilient control for linear discrete-time multi-agent networks under
[9] K. W. Suh, Modeling, Analysis and Control of Fuel Cell Hybrid Power unconfined cyber-attacks,” IEEE Trans. Circuits Syst. I, Reg. Papers,
Systems. Ann Arbor, MI, USA: Univ. Michigan, 2006. vol. 68, no. 2, pp. 776–785, Feb. 2021.

Authorized licensed use limited to: GUANGZHOU UNIVERSITY. Downloaded on February 02,2025 at 10:16:59 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

14 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS

[31] H. N. Vu, D. Truong Le Tri, H. L. Nguyen, Y. Kim, and S. Yu, “Mul- Shihong Ding (Senior Member, IEEE) was born
tifunctional bypass valve for water management and surge protection in Anhui, China, in 1983. He received the B.E.
in a proton-exchange membrane fuel cell supply-air system,” Energy, degree in mathematics from Anhui Normal Uni-
vol. 278, Sep. 2023, Art. no. 127696. versity, Wuhu, China, in 2004, and the M.S. and
[32] K. Mei, S. Ding, and X. Yu, “A generalized supertwisting algorithm,” Ph.D. degrees in automatic control from South-
IEEE Trans. Cybern., vol. 53, no. 6, pp. 3951–3960, Jun. 2023. east University, Nanjing, China, in 2007 and 2010,
[33] J. Yu, X. Dong, Q. Li, J. Lu, and Z. Ren, “Adaptive practical optimal respectively.
time-varying formation tracking control for disturbed high-order multi- During his graduate studies, he visited The Uni-
agent systems,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 69, no. 6, versity of Texas at San Antonio, San Antonio, TX,
pp. 2567–2578, Jun. 2022. USA, from August 2008 to August 2009. After
[34] H. Wang, K. Xu, and J. Qiu, “Event-triggered adaptive fuzzy fixed-time graduation, he held a research fellowship with the
tracking control for a class of nonstrict-feedback nonlinear systems,” University of Western Sydney, Penrith, NSW, Australia, for one year.
IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 68, no. 7, pp. 3058–3068, He also visited Yeungnam University, Gyeongsan, South Korea, from July
Jul. 2021. 2018 to August 2018, and RMIT University, Melbourne, VIC, Australia,
[35] X. Wang, Y. Zhou, B. Luo, Y. Li, and T. Huang, “Event-triggered from December 2019 to February 2020. Since June 2010, he has been with
neuro-adaptive fixed-time control for nonlinear switched and constrained the School of Electrical and Information Engineering, Jiangsu University,
systems: An initial condition-independent method,” IEEE Trans. Circuits Zhenjiang, China, where he is currently a Full Professor. His research interests
Syst. I, Reg. Papers, vol. 71, no. 5, pp. 2229–2239, May 2024. include sliding-mode control and finitetime stability.
[36] K. B. Ariyur and M. Krstic, Real-Time Optimization By Extremum- Prof. Ding currently serves as a Subject Editor for Nonlinear Dynamics and
seeking Control. Hoboken, NJ, USA: Wiley, 2003. an Associate Editor for IEEE ACCESS.
[37] Z. Wang, X. Zhou, and J. Wang, “Extremum-seeking-based adaptive
model-free control and its application to automated vehicle path track-
ing,” IEEE/ASME Trans. Mechatronics, vol. 27, no. 5, pp. 3874–3884,
Oct. 2022.
[38] J. Xu and Z. Jin, “Extremum-seeking for PWM-controlled double- Jing Zhao (Member, IEEE) received the Ph.D.
integrator system via nonlinear small-gain approach,” IEEE Trans. degree in electromechanical engineering from the
Circuits Syst. I, Reg. Papers, vol. 70, no. 12, pp. 5433–5441, Dec. 2023. University of Macau, Macau, China, in 2016.
He is currently working with the School of
[39] B. Hunnekens, A. D. Dino, N. van de Wouw, N. van Dijk, and
Mechanical Engineering and Automation, Northeast-
H. Nijmeijer, “Extremum-seeking control for the adaptive design of
ern University. His research interests include vehicle
variable gain controllers,” IEEE Trans. Control Syst. Technol., vol. 23,
dynamics and control, mechanism and machine the-
no. 3, pp. 1041–1051, May 2015.
ory, fluid mechanics and finite element analysis.
[40] M. Krstić, “Performance improvement and limitations in extremum
seeking control,” Syst. Control Lett., vol. 39, no. 5, pp. 313–326,
Apr. 2000.
[41] G. T. Gilbert, “Positive definite matrices and Sylvester’s criterion,” Amer.
Math. Monthly, vol. 98, no. 1, pp. 44–46, 1991.
[42] H. K. Khalil, Control of Nonlinear Systems. New York, NY, USA:
Prentice-Hall, 2002.
Jinwu Gao received the B.Eng. degree from the
[43] H. Obeid, S. Laghrouche, L. Fridman, Y. Chitour, and M. Harmouche, Department of Automation Measurement and Con-
“Barrier function-based adaptive super-twisting controller,” IEEE Trans. trol Engineering and the Ph.D. degree from the
Autom. Control, vol. 65, no. 11, pp. 4928–4933, Nov. 2020. Department of Control Science and Engineering,
[44] J. Gao, M. Li, Y. Hu, H. Chen, and Y. Ma, “Challenges and developments Harbin Institute of Technology, Harbin, China, in
of automotive fuel cell hybrid power system and control,” Sci. China 2005 and 2012, respectively.
Inf. Sci., vol. 62, no. 5, pp. 1–25, May 2019. From 2012 to 2014, he was an Assistant Pro-
fessor with Sun Yat-sen University, Guangzhou,
China. In July 2014, he held a postdoctoral posi-
tion with the Department of Engineering and
Applied Science, Sophia University, Tokyo, Japan.
From 2016 to 2020, he was an Associate Professor with Jilin University,
Lin Chen received the B.E. degree from the College Changchun, China, where he has been a Professor since September 2020.
of Communication Engineering, Jilin University, His research interests include control theory and application in automotive
Changchun, China, in 2021, where he is currently powertrain.
pursuing the Ph.D. degree with the Department of
Control Science and Engineering.
His current research interests include system con-
trol and efficiency optimization for proton exchange
membrane fuel cells.
Hong Chen (Fellow, IEEE) received the B.S. and
M.S. degrees in process control from Zhejiang
University, Hangzhou, China, in 1983 and 1986,
respectively, and the Ph.D. degree in system dynam-
ics and control engineering from the University of
Stuttgart, Stuttgart, Germany, in 1997.
In 1986, she joined Jilin University of Technology,
Jinfa Liu received the B.S. degree in automation Changchun, China. From 1993 to 1997, she was
from Jilin University, Changchun, China, in 2022, a Wissenschaftlicher Mitarbeiter with the Institut
where he is currently pursuing the M.S. degree with fuer Systemdynamik und Regelungstechnik, Uni-
the Department of Control Science and Engineering. versity of Stuttgart. Since 1999, she has been a
His current research interests include modeling and Professor with Jilin University and hereafter a Tang Aoqing Professor.
optimization of fuel cell electrolysers. From 2015 to 2019, she was the Director of the State Key Laboratory of Auto-
motive Simulation and Control. She is currently a Distinguished Professor
of Tongji University, Shanghai, China. Her current research interests include
model predictive control, nonlinear control, and applications in mechatronic
systems focusing on automotive systems.

Authorized licensed use limited to: GUANGZHOU UNIVERSITY. Downloaded on February 02,2025 at 10:16:59 UTC from IEEE Xplore. Restrictions apply.