betaFeedback

WorksEngineering

Model-Free Detection and Accommodation of Sensor Faults for a PEM Electrolyzer

Abstract

We investigate the detection and accommodation of sensor faults for a proton exchange membrane electrolyzer coupled to a DC/DC converter powered by renewable energy sources. The proposed method for detecting and accommodating the sensor fault is model-free and is based on the concept of ultra-local model that is becoming classic in control engineering. The existing literature on active control tolerant to sensor fault dedicated to this question shows that no previous work has addressed this topic. Our approach mitigates the effect of sensor fault on closed-loop behavior and guarantees the stability and performance of the overall system. Numerical simulations under variations in renewable energy sources validate our approach.

Introduction

Green hydrogen is playing an key part in reducing carbon emissions in the industrial and transport sectors. This hydrogen is mainly produced using water electrolysers powered by renewable energy sources. In this framework, Proton Exchange Membrane Water Electrolyzers (PEMWE) technology is widely used for hydrogen production due to its fast response to renewable energy sources variations and its ability to support high current densities of up to 2 A/cm22\ \mathrm{A/cm^2} [1].

Due to its complexity, the PEMWE system is subject to several types of faults: those occurring at the stack or cell level, such as membrane drying, crossover, membrane degradation, . . . , as well as those appearing at the system level, including balance of plant (leaks in pumps, valves, sensors, etc.). In recent years, many studies have been devoted to the diagnosis of this system. A model-based diagnosis approach via bond graph is proposed in [2]. The detection and isolation of several stack faults as membrane flooding and drying, crossover, . . . are made by evaluating residuals. A signal-based approach is proposed in [3] to detect stack short circuits and water shortages and to detect pump fault in [4]. However, these methods consider the open loop operation of the system.

A PEMWE operates at a voltage significantly lower than that provided by Renewable Energy Sources (RES), thereby necessitating a DC/DC buck converter as an interface [5], [6]. An Active Fault-Tolerant Control (AFTC) strategy has been

proposed for a three-phase stacked interleaved buck converter supplying a PEM electrolyzer [7], where a gate fault in one phase is considered. The proposed strategy enables continued operation under a degraded two-phase mode. Nevertheless, AFTC strategies addressing sensor faults accommodation for PEMWE systems remain largely unexplored and is the purpose of this work.

This paper proposes an active sensor fault-tolerant control strategy for a PEMWE system coupled with DC/DC converter supplied by RES. The proposed approach ensures closed-loop operation under RES voltage variations and is model-free, i.e. it does not rely on a mathematical modeling of the system for fault detection and accommodation.

This paper is organized as follows. Section II describes the PEMWE and DC/DC converter models which is used as a test bench for validation and states the problem of sensor AFTC. Section III presents the proposed model-free sensor fault accommodation approach, accounting for variations in the input voltage. Simulation results are provided in Section IV, and conclusions are drawn in Section V.

PEMWE Modeling and Control Objective

Modeling and control objective

The PEMWE model used in this paper is the one given in [8] which has been identified using Electrochemical Impedance Spectroscopy (EIS) method detailed in [9], [10]. The best obtained model in [8] is the one of order 6 and is given by (1).

In order to obtain a model for the whole PEMWE + SIBC converter, we need to express the transfer function in (1) with passive electrical components. This can be achieved using the residue theorem [11] as follows

G(s)=∑h=16RhChRhs+1+R0,(1)G(s) = \sum_{h=1}^{6} \frac{R_h}{C_h R_h s + 1} + R_0, \tag*{(1)}

where

R6=100.06 Ω, R5=13.588 Ω, R4=1.4068 Ω,R_6 = 100.06\ \Omega,\ R_5 = 13.588\ \Omega,\ R_4 = 1.4068\ \Omega,

R3=0.20564 Ω, R2=0.040826 Ω, R1=0.0085339 Ω,R_3 = 0.20564\ \Omega,\ R_2 = 0.040826\ \Omega,\ R_1 = 0.0085339\ \Omega,

C6=1.0361×10−6 F, C5=9.4843×10−5 F,C_6 = 1.0361 \times10^{-6}\ \mathrm{F},\ C_5 = 9.4843 \times10^{-5}\ \mathrm{F},

C4=0.0086629 F, C3=0.72622 F, C2=27.953 F,C_4 = 0.0086629\ \mathrm{F},\ C_3 = 0.72622\ \mathrm{F},\ C_2 = 27.953\ \mathrm{F},

C1=901.43 F, R0=0.027366 Ω.C_1 = 901.43\ \mathrm{F},\ R_0 = 0.027366\ \Omega.

So the coupling of the PEMWE model given by (2) with the SIBC converter is represented by Figure 1 where VRESV_{RES} is the voltage supplied by the RES, uu is the control input and

Figure 1. Stacked interleaved DC-DC buck converter with pulse width modulation (PWM) and electrical model of PEMWE

G(s)=0.02737s6+403s5+4.667e05s4+5.015e07s3+5.068e08s2+6.438e08s+1.009e08s6+1.051e04s5+8.42e06s4+6.784e08s3+4.788e09s2+4.214e09s+4.684e08(2)G(s)=\frac{0.02737s^{6}+403s^{5}+4.667e05s^{4}+5.015e07s^{3}+5.068e08s^{2}+6.438e08s+1.009e08}{s^{6}+1.051e04s^{5}+8.42e06s^{4}+6.784e08s^{3}+4.788e09s^{2}+4.214e09s+4.684e08} \tag*{(2)}

the SIBC electrical components are given by: Lp=Ls=426×10−6 HL_p=L_s=426\times10^{-6}\,\mathrm{H}, Rℓp=Rℓs=0.06 ΩR_{\ell p}=R_{\ell s}=0.06\,\Omega, Cp=10−4 FC_p=10^{-4}\,\mathrm{F} and Cs=10−5 FC_s=10^{-5}\,\mathrm{F}.

Using the average state-space approach proposed in [12] and detailed in [13], a state-space model for the system given in Figure 1 is derived in [14] in the case of the number of parallel RCRC cells is equal to 2 instead of 6 in (2). So using straightforward developments, the state-space model in [14] leads to the following one

x˙(t)=Ax(t)+BVRES(t)u(t)(3)\dot{x}(t)=Ax(t)+BV_{\mathrm{RES}}(t)u(t) \tag*{(3)}
y(t)=Cx(t)(4)y(t)=Cx(t) \tag*{(4)}

where u(t)u(t), x(t)x(t) and y(t)y(t) are the control law, the state vector and the regulated output respectively, where

x(t)T=[ipisVvCsvC1⋯vC6](5)x(t)^{T}=\begin{bmatrix}i_p&i_s&V&v_{C_s}&v_{C_1}&\cdots&v_{C_6}\end{bmatrix} \tag*{(5)}
y(t)=I(t).(6)y(t)=I(t). \tag*{(6)}

The state-space matrices are given by

A=[−RℓpLp0−1Lp001×60−RℓsLs−1Ls−1Ls01×61Cp1Cs1R0Cp0−11×6R0Cp01Cs0001×606×106×1A006×1A]A=\begin{bmatrix} -\frac{R_{\ell p}}{L_p} & 0 & -\frac{1}{L_p} & 0 & 0_{1\times6}\\ 0 & -\frac{R_{\ell s}}{L_s} & -\frac{1}{L_s} & -\frac{1}{L_s} & 0_{1\times6}\\ \frac{1}{C_p} & \frac{1}{C_s} & \frac{1}{R_0C_p} & 0 & -\frac{1_{1\times6}}{R_0C_p}\\ 0 & \frac{1}{C_s} & 0 & 0 & 0_{1\times6}\\ 0_{6\times1} & 0_{6\times1} & A_0 & 0_{6\times1} & A \end{bmatrix}
B=[1Lp1Ls08×1],C=[001R00−11×6R0]B=\begin{bmatrix} \frac{1}{L_p}\\ \frac{1}{L_s}\\ 0_{8\times1} \end{bmatrix}, \qquad C=\begin{bmatrix} 0 & 0 & \frac{1}{R_0} & 0 & -\frac{1_{1\times6}}{R_0} \end{bmatrix}

where A∈R6×6A\in\mathbb{R}^{6\times6} and A0∈R6×1A_0\in\mathbb{R}^{6\times1} are given by (for h=1,…,6h=1,\ldots,6 and q=1,…,6q=1,\ldots,6)

A(h,h)=−1RhCh+−1R0Ch,A(h,h)=-\frac{1}{R_{h}C_{h}}+\frac{-1}{R_0C_{h}},
A(h,q)=−1R0Ch,h≠q,A(h,q)=\frac{-1}{R_0C_{h}},\qquad h\ne q,
A0(h)=1R0Ch.A_0(h)=\frac{1}{R_0C_{h}}.

In this paper, the RES supplied voltage VRES(t)V_{\mathrm{RES}}(t) is known and bounded as

VRES(t)∈[VRESm VRESM](7)V_{\mathrm{RES}}(t)\in[V_{\mathrm{RESm}}\ V_{\mathrm{RESM}}] \tag*{(7)}

where VRESm=25 VV_{\mathrm{RESm}}=25\,\mathrm{V} and VRESM=55 VV_{\mathrm{RESM}}=55\,\mathrm{V}.

Based on Faraday’s law [15], the flow rate hydrogen m˙H2\dot{m}_{\mathrm{H}_2} produced by PEMWE is given by

m˙H2=NcI2FηH2(8)\dot{m}_{\mathrm{H}_2}=\frac{N_cI}{2F}\eta_{\mathrm{H}_2} \tag*{(8)}

where Nc=3N_c=3 is the number of cells in the stack, F=96485F=96485 is Faraday’s constant, and ηH2\eta_{\mathrm{H}_2} is Faraday’s efficiency coefficient taken as 0.97. So, considering the hydrogen flow rate as the control objective is like controlling the current II at the desired setpoint I∗I^*.

Problem statement

The measured current ymy_m is expressed as

ym=Im=y+f+ω(9)y_m = I_m = y + f + \omega \tag*{(9)}

where yy is defined in (6), ff is a sensor fault, and ω\omega is a zero-mean white Gaussian noise.

If the sensor fault ff is not taken into account in the synthesis of the control law, then it is ImI_m and not II that converges to the desired setpoint value I⋆I^\star. So an AFTC is therefore necessary to compensate the effect of ff on the closed-loop behavior in order to ensure that II converges to I⋆I^\star despite f≠0f \ne0 and the variations of VRES(t)V_{\mathrm{RES}}(t).

This problem is solved in Sect. III by employing our model-free viewpoint on fault detection, estimation, and accommodation.

Remark 1: It is important to note that the proposed approach detailed in Sect. III does not use Eq. (3), which will be used in Sect. IV to emulate the PEMWE + SIBC behavior. □

Model-free sensor fault accommodation for PEMWE

Before presenting the AFTC procedure, we briefly describe the model-free control design that is used in this paper.

Model-free control

Model-free control is based on an ultra-local model given by [16, 17]

y˙m=F+αu(10)\dot{y}_m = \mathcal{F} + \alpha u \tag*{(10)}

where F\mathcal{F} corresponds to the unknown part, uu and ymy_m are the control input and measured output, α\alpha is an appropriate constant coefficient. An estimation of F\mathcal{F} reads [16]

F^(t)=−6τ3∫t−τt[(τ−2σ)ym(σ)+ασ(τ−σ)u(σ)]dσ\widehat{\mathcal{F}}(t) = -\frac{6}{\tau^3}\int_{t-\tau}^{t} \left[(\tau-2\sigma)y_m(\sigma)+\alpha\sigma(\tau-\sigma)u(\sigma)\right]\mathrm{d}\sigma

The corresponding closed-loop control law, or intelligent proportional controller (iP), reads [16]

u=−F^−y˙⋆+Kpeα(11)u = -\frac{\widehat{\mathcal{F}}-\dot{y}^{\star}+K_p e}{\alpha} \tag*{(11)}

where y⋆=I⋆y^\star=I^\star is the reference trajectory, e=ym−y⋆e=y_m-y^\star is the tracking error, while KpK_p is a tuning gain.

In the presence of a sensor fault f≠0f\ne0, the tracking error ee converges to 0 but yy does not track asymptotically y⋆y^\star since ym≠yy_m\ne y.

Model-free active fault tolerant control

Residual generation

We consider the following residual

r(t)=ym(t)−β(t)y^m(t)(12)r(t)=y_m(t)-\beta(t)\widehat{y}_m(t) \tag*{(12)}

where y^m\widehat{y}_m is given by

y^m(t)=∫0t(F^(σ)+αu(σ))dσ+y^m(0)(13)\widehat{y}_m(t)=\int_{0}^{t}\left(\widehat{\mathcal{F}}(\sigma)+\alpha u(\sigma)\right)\mathrm{d}\sigma+\widehat{y}_m(0) \tag*{(13)}

and β(t)\beta(t) is a parameter to be determined. For sensor fault detection and estimation purposes the residual in Eq. (12) must satisfy the following conditions

  • if f(t)=0f(t)=0, then r(t)=0r(t)=0, ∀VRES(t)\forall V_{\mathrm{RES}}(t),

  • if f(t)≠0f(t)\ne0, then r(t)≠0r(t)\ne0, ∀VRES(t)\forall V_{\mathrm{RES}}(t).

So, when f(t)=0f(t)=0, the residual r(t)r(t) is equal to 0, if

β(t)=ym(t)y^m(t)=ym(t)∫0t(F^(σ)+αu(σ))dσ+y^m(0)(14)\beta(t)=\frac{y_m(t)}{\widehat{y}_m(t)}=\frac{y_m(t)}{\int_{0}^{t}\left(\widehat{\mathcal{F}}(\sigma)+\alpha u(\sigma)\right)\mathrm{d}\sigma+\widehat{y}_m(0)} \tag*{(14)}

where the steady-state values of the ymy_m and y^m\widehat{y}_m are used. It has been shown [18, 19] that β(t)\beta(t) in Eq. (12) is constant for Linear Time Invariant (LTI) systems. But the PEMWE system in (4) is not LTI due to the variations of VRESV_{\mathrm{RES}}. Those variations are influencing y^m(t)\widehat{y}_m(t) and β(t)\beta(t). It follows that β(t)\beta(t) should be expressed as depending on VRES(t)V_{\mathrm{RES}}(t). Since the iP controller guarantees local stability, the steady-state is still achieved and the control input u(t)u(t) in steady-state can be expressed as follows

u(t)=c(t)y(t)=c(t)ym(t)(15)u(t)=c(t)y(t)=c(t)y_m(t) \tag*{(15)}

To approximate the integral in Eq. (14), the well-known rectangle method is used, where TeT_e is the sampling time:

∫0tq(σ)dσ≃∑i=1kq(i)Te\int_{0}^{t}q(\sigma)\mathrm{d}\sigma\simeq\sum_{i=1}^{k}q(i)T_e

and

q˙(t)≃δq(i)=q(i)−q(i−1)Te\dot{q}(t)\simeq\delta_q(i)=\frac{q(i)-q(i-1)}{T_e}

After straightforward computations, y^m\widehat{y}_m is expressed as follows (see [20] for more details):

y^m(kTe)=ym(kTe)+αTeu(kTe)−ym(0)+y^m(0)(16)\widehat{y}_m(kT_e)=y_m(kT_e)+\alpha T_e u(kT_e)-y_m(0)+\widehat{y}_m(0) \tag*{(16)}

where F^(kTe)\widehat{\mathcal{F}}(kT_e) is given by (see Eq. (10))

F^(kTe)=δym(kTe)−αu((k−1)Te)(17)\widehat{\mathcal{F}}(kT_e)=\delta_{y_m}(kT_e)-\alpha u((k-1)T_e) \tag*{(17)}

Insert Eq. (15) in Eq. (16), assume that ym(0)≃y^m(0)y_m(0)\simeq\widehat{y}_m(0), then

β(kTe)=ym(kTe)ym(kTe)(1+Teαc(kTe))=11+Teαc(kTe)(18)\beta(kT_e)=\frac{y_m(kT_e)}{y_m(kT_e)\left(1+T_e\alpha c(kT_e)\right)}=\frac{1}{1+T_e\alpha c(kT_e)} \tag*{(18)}

Let VRESiV_{\mathrm{RES}_i} and cic_i be the values of VRES(kiTe)V_{\mathrm{RES}}(k_iT_e) and c(kiTe)c(k_iT_e), where kiTek_iT_e is a given time-instant in the steady-state behavior, where k>kik>k_i.

Since the closed-loop stability and steady-state are achieved, x˙=0\dot{x}=0 in Eq. (3), c(kTe)c(kT_e) in Eq. (15) reads

c(kTe)=ci(1+δc(kTe))(19)c(kT_e)=c_i(1+\delta_c(kT_e)) \tag*{(19)}

From

VRES(kTe)=VRESi(1+δVRES(kTe))(20)V_{\mathrm{RES}}(kT_e)=V_{\mathrm{RES}_i}(1+\delta_{V_{\mathrm{RES}}}(kT_e)) \tag*{(20)}

we obtain

δc(kTe)=−δVRES(kTe)1+δVRES(kTe)(21)\delta_c(kT_e)=\frac{-\delta_{V_{\mathrm{RES}}}(kT_e)}{1+\delta_{V_{\mathrm{RES}}}(kT_e)} \tag*{(21)}

Via Eqs. (19), (20), (21) and (16), β(kTe)\beta(kT_e) becomes

β(kTe)=11+VRESi(1−βi)VRES(kTe)βi(22)\beta(kT_e)=\frac{1}{1+\frac{V_{\mathrm{RES}_i}(1-\beta_i)}{V_{\mathrm{RES}}(kT_e)\beta_i}} \tag*{(22)}

where βi=β(kiTe)\beta_i=\beta(k_iT_e). 2) Sensor fault estimation and accommodation: If f≠0f \ne0 in Eq. (9), we obtain

u(kTe)=c(kTe)y(kTe)=c(kTe)(ym(kTe)−f(kTe)−ω(kTe))(23)u(kT_e)=c(kT_e)y(kT_e) =c(kT_e)\left(y_m(kT_e)-f(kT_e)-\omega(kT_e)\right) \tag*{(23)}
y^m(kTe)=ym(kTe)+αTec(kTe)(ym(kTe)−f^(kTe))(24)\hat{y}_m(kT_e)=y_m(kT_e)+\alpha T_e c(kT_e)\left(y_m(kT_e)-\hat{f}(kT_e)\right) \tag*{(24)}

where the initial condition is chosen by y^m(0)≃ym(0)\hat{y}_m(0)\simeq y_m(0), and f^\hat{f} is the estimation of the sensor fault ff to be determined. Notice that f^\hat{f} is used in Eq. (24) in place of ff since y^m\hat{y}_m does not depend explicitly of the noise ω\omega.

Via Eqs. (22), (24), and (12), we get

r(kTe)=ym(kTe)−11+Teαc(kTe)×(ym(kTe)(1+αTec(kTe))−αTec(kTe)f^(kTe))=αTec(kTe)f^(kTe)1+αTec(kTe)=f^(kTe)(1−11+αTec(kTe))=f^(kTe)(1−β(kTe))(25)\begin{aligned} r(kT_e)&=y_m(kT_e)-\frac{1}{1+T_e\alpha c(kT_e)}\times\\ &\quad\left(y_m(kT_e)(1+\alpha T_e c(kT_e))-\alpha T_e c(kT_e)\hat{f}(kT_e)\right)\\ &=\frac{\alpha T_e c(kT_e)\hat{f}(kT_e)}{1+\alpha T_e c(kT_e)}\\ &=\hat{f}(kT_e)\left(1-\frac{1}{1+\alpha T_e c(kT_e)}\right)\\ &=\hat{f}(kT_e)(1-\beta(kT_e)) \tag*{(25)} \end{aligned}

and the best estimation of sensor fault f(kTe)f(kT_e) is given by

f^(kTe)=r(kTe)1−β(kTe)(26)\hat{f}(kT_e)=\frac{r(kT_e)}{1-\beta(kT_e)} \tag*{(26)}

where β(kTe)≠1\beta(kT_e)\ne1 (see Eq. (18)).

The threshold of the residual r(kTe)r(kT_e) is called ρ(kTe)\rho(kT_e). It is given by

ρ(kTe)=f~∣1−β(kTe)∣(27)\rho(kT_e)=\tilde{f}|1-\beta(kT_e)| \tag*{(27)}

where f~\tilde{f} is the minimal absolute value of the sensor fault f^(kTe)\hat{f}(kT_e) to be detected and estimated. Once the sensor fault is detected and estimated, i.e, ∣r(t)∣>∣ρ(t)∣|r(t)|>|\rho(t)|, the sensor fault accommodation is made by adapting the tracking error defined in (9) as follows

e(kTe)=ym(kTe)−y∗(kTe)−f^(kTe)(28)e(kT_e)=y_m(kT_e)-y^*(kT_e)-\hat{f}(kT_e) \tag*{(28)}

Simulations results

Two cases are considered: the 1{sup}st one with constant VRES=40V_{RES}=40 V corresponding to Figure 3 and Figure 4 and the 2{sup}nd one with variable VRESV_{RES} which is shown in Figure 2 corresponding to Figure 5 and Figure 6. In these two cases, there is no sensor fault accommodation procedure in Figure 3 and Figure 5, while the AFTC is applied in Figure 4 and Figure 6. Transient behaviors can be distinguished from steady-state behaviors: colored areas correspond to transient behaviors, while uncolored areas are associated with steady-state behaviors.

VRES variations

Figure 2. VRESV_{RES} variations

Four-panel plots showing trajectory tracking, residual, estimated sensor fault, and control input without AFTC

Figure 3. (a) Trajectory tracking without AFTC (b) Residual rr (c) Estimated sensor fault f^\hat{f} (d) Control input uu. Constant VRESV_{RES} without AFTC

Four-panel plots showing trajectory tracking, residual, estimated sensor fault, and control input with AFTC

Figure 4. (a) Trajectory tracking with AFTC (b) Residual rr (c) Estimated sensor fault f^\hat{f} (d) Control input uu. Constant VRESV_{RES} with AFTC

Trajectory tracking without AFTC

(a) (a) Trajectory tracking without AFTC

Residual $r$

(b) (b) Residual rr

Estimated sensor fault $\hat{f}_b$

(c) (c) Estimated sensor fault f^b\hat{f}_b

Control input $u$

(d) (d) Control input uu

Figure 6. Variable VRESV_{\mathrm{RES}} with AFTC

Trajectory tracking without AFTC

(a) (a) Trajectory tracking without AFTC

Residual $r$

(b) (b) Residual rr

Estimated sensor fault $\hat{f}_b$

(c) (c) Estimated sensor fault f^b\hat{f}_b

Control input $u$

(d) (d) Control input uu

Figure 5. Variable VRESV_{\mathrm{RES}} without AFTC

In Figure 3 to Figure 6, ρ+(kTe)=ρ(kTe)\rho^+(kT_e)=\rho(kT_e) and ρ−(kTe)=−ρ(kTe)\rho^-(kT_e)=-\rho(kT_e) where f~=0.5\tilde{f}=0.5 A (see Eq. (27)). In the measurement Eq. (9), ω\omega is a zero-mean white Gaussian noise with standard deviation σω=0.0231\sigma_\omega=0.0231. The value of the iPiP controller parameters are α=9750\alpha=9750 and Kp=100K_p=100.

For all treated cases, we can see that the iPiP controller ensures closed-loop stability and asymptotic trajectory tracking. So, the AFTC procedure can be applied on the steady-state behavior. For the 1{sup}st case, the sensor fault f=2f=2 A appears at time t=30t=30 s (see Figure 3c and Figure 4c) and the threshold ρ\rho is constant (see Figure 3b and Figure 4b), while the occurrence of the sensor fault f=−4f=-4 A appears at time t=60t=60 s for the 2{sup}nd case (see Figure 5c and Figure 6c) and the threshold ρ\rho is slightly variable (see Figure 5b and Figure 6b).

For the case where there is no accommodation procedure, the following comments can be made on Figure 3 and Figure 5:

  • the residual rr exceeds the threshold ρ\rho at a time-instant very close from the occurrence of the sensor fault, allowing the fault detection and estimation (see Figure 3b and Figure 5b);

  • the sensor fault is well estimated (see Figure 3c and Figure 5c);

  • after the fault occurrence, the measurement ymy_m follows the desired trajectory y∗y^* instead of the regulated output yy (see Figure 3a and Figure 5a).

For the case where the AFTC procedure is applied, the following comments can be made on Figure 4 and Figure 6:

  • the residual rr exceeds the threshold ρ\rho at a time-instant very close from the occurrence of the sensor fault, allowing the fault detection and estimation (see Figure 4b and Figure 6b). At this time-instant, the AFTC procedure is triggered;

  • the sensor fault is well estimated (see Figure 4c and Figure 6c);

  • once the AFTC procedure is triggered, the regulated output yy follows the desired trajectory y∗y^*, which means that the AFTC is effective (see Figure 4a and Figure 6a). Notice that, the AFTC is not triggered in the transient behavior, so the measurement ymy_m follows the desired trajectory y∗y^* instead of the regulated output yy;

  • by comparing Figure 3d and Figure 4d on the one hand, and Figure 5d and Figure 6d on the other, we can see how the AFTC procedure has modified the control input uu.

Conclusions

Detection and accommodation of current sensor fault for the PEMWE were discussed. This measurement is essential for controlling hydrogen production via a water electrolyzer. The method developed does not rely on an analytical model of the process: the detection and compensation of sensor faults are model-free. The simulation results confirm that the actual current of the electrolyzer is regulated to the desired current despite significant variations in the voltage supplying the process in the presence of a sensor fault.

The fact that our approach is model-free and based on an ultra-local model makes it a good candidate for real-time application on a test bench, with low computational burden. The implementation of this method on a real test bench is being considered as part of future work.

References

  1. [1]D. Falcão and A. Pinto, “A review on PEM electrolyzer modelling: guidelines for beginners,” J. Cleaner Production, vol. 261, p. ID 121184, 2020.DOI
  2. [2]S. Sood, O. Prakash, J.-Y. Dieulot, M. Boukerdja, B. Ould-Bouamaama, and M. Bressel, “Robust diagnosis of PEM electrolysers using LFT bond graph,” International Journal of Hydrogen Energy, vol. 47, no. 80, pp. 33 938–33 954, 2022.
  3. [3]D. Shen, Y. Dong, Z. Lei, F. Yang, L. Cheng, and J. Jing, “Fault diagnosis of cells in pem electrolyzer stack under fluctuating power source,” International Journal of Hydrogen Energy, vol. 97, pp. 350–361, 2025.
  4. [4]F. Aubras, C. Damour, M. Benne, S. Boulevard, M. Bessafi, B. Grondin-Perez, A. J.-J. Kadjo, and J. Deseure, “A non-intrusive signal-based fault diagnosis method for proton exchange membrane water electrolyzer using empirical mode decomposition,” Energies, vol. 14, no. 15, p. 4458, 2021.
  5. [5]V. Guida, D. Guilbert, G. Vitale, and B. Douine, “Design and realization of a stacked interleaved DC-DC step-down converter for PEM water electrolysis with improved current control,” Fuel Cells, vol. 20, pp. 307–315, 2020.DOI
  6. [6]B. Yodwong, D. Guilbert, W. Kaewmanee, M. Phattanasak, M. Hinaje, and G. Vitale, “Improved sliding mode-based controller of a high voltage ratio DC-DC converter for electrolyzers supplied by renewable energy,” IEEE Trans. Industrial Electronics, vol. 71, pp. 8831–8840, 2024.DOI
  7. [7]X. Guo, S. Zhang, Z. Liu, L. Sun, Z. Lu, C. Hua, and J. M. Guerrero, “A new multi-mode fault-tolerant operation control strategy of multiphase stacked interleaved buck converter for green hydrogen production,” International Journal of Hydrogen Energy, vol. 47, no. 71, pp. 30 359–30 370, 2022.
  8. [8]M. Zasadzinski, M. Ait Ziane, and M. Alma, “Experimental identification approaches for a PEM water electrolyzer,” in Proc. IEEE Int. Conf. Control, Decision and Information Technologies, Split, Croatia, 2025.DOI
  9. [9]T. Vidaković-Koch, T. Miličić, L. Živković, H. Seng Chan, U. Krewer, and M. Petkovska, “Nonlinear frequency response analysis: a recent review and perspectives,” Current Opinion in Electrochemistry, vol. 30, p. ID 100851, 2021.
  10. [10]T. Miličić, M. Sivasankaran, C. Blümner, A. Sorrentino, and T. Vidaković-Koch, “Pulsed electrolysis — explained,” Faraday Discussions, vol. 246, pp. 179–197, 2023.DOI
  11. [11]K. Ogata, Modern Control Engineering. Englewood Cliffs, New Jersey: Prentice Hall, 1970.
  12. [12]R. Middlebrook and S. Cuk, “A general unified approach to modelling switching-converter power stages,” in Proc. IEEE Power Electronics Specialists Conf., Cleveland, USA, 1976.
  13. [13]N. Mohan, T. Undeland, and W. Robbins, Power Electronics, 2nd ed. New York: Wiley, 1995.
  14. [14]R. Maamouri, D. Guilbert, M. Zasadzinski, and H. Rafaralahy, “Proton exchange membrane water electrolysis: modeling for hydrogen flow rate control,” Int. J. of Hydrogen Energy, vol. 46, pp. 7676–7700, 2021.DOI
  15. [15]M. Ait Ziane, M. Zasadzinski, A. Bouramdane, E. Pahon, and H. Rafaralahy, “Experimental control of a PEM water electrolyzer: investigation of renewable energy source framework,” in Proc. IEEE Int. Conf. Control, Decision and Information Technologies, Split, Croatia, 2025.
  16. [16]M. Fliess and C. Join, “Model-free control,” Int. J. Control, vol. 86, pp. 2228–2252, 2013.
  17. [17]———, “An alternative to proportional-integral and proportional-integral-derivative regulators: Intelligent proportional-derivative regulators,” Int. J. Robust & Nonlinear Contr., vol. 32, pp. 9512–9524, 2022.DOI
  18. [18]M. Ait Ziane, C. Join, and M. Zasadzinski, “Sensor fault accommodation: a model-free framework,” Int. J. Control, 2025, https://doi.org/10.1080/00207179.2025.2482182.
  19. [19]M. Ait Ziane, M. Zasadzinski, C. Join, and M. Péra, “Model-free active sensor fault tolerant control for systems with linear static characteristic,” Int. J. of Dynamics and Control, vol. xx, pp. xx–xx, 2024, submitted.DOI
  20. [20]———, “Model-free active sensor fault tolerant control for systems with linear static characteristic,” Int. J. of Dynamics and Control, vol. 13, p. ID 267, 2025.DOI

Paper details

vlnc:W_28e84zqu
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents