You are on page 1of 17

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 6 ( 2 0 1 1 ) 7 2 1 2 e7 2 2 8

Available at www.sciencedirect.com

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

Review

A Review on solid oxide fuel cell models


ra a, N. Steiner b, D. Marra c, M. Sorrentino c, C. Pianese c, K. Wang a,*, D. Hissel a, M.C. Pe M. Monteverde d, P. Cardone d, J. Saarinen e
a

, FEMTO-ST (UMR CNRS 6174), FCLAB, 90000 Belfort, France University of Franche-Comte EIFER, European Institute For Energy Research, Emmy-Noether Strasse 11, 76131 Karlsruhe, Germany c Department of Industrial Engineering, University of Salerno, 84084 Fisciano, SA, Italy d Faculty of Engineering, University of Genoa, 16145 Genoa, Italy e VTT Technical Research Centre of Finland, Biologinkuja, P.O. Box 1000, FI-02044 VTT, Finland
b

article info
Article history: Received 10 December 2010 Received in revised form 6 March 2011 Accepted 10 March 2011 Available online 13 April 2011 Keywords: SOFC Modelling Articial intelligent Neural network Electrochemical impedance spectroscopy Model-based diagnosis

abstract
Since the model plays an important role in diagnosing solid oxide fuel cell (SOFC) system, this paper proposes a review of existing SOFC models for model-based diagnosis of SOFC stack and system. Three categories of modelling based on the white-, the black- and the grey-box approaches are introduced. The white-box model includes two types, i.e. physical model and equivalent circuit model based on EIS technique. The black-box model is based on articial intelligence and its realisation relies mainly on experimental data. The greybox model is more exible: it is a physical representation but with some parts being modelled empirically. Validation of models is discussed and a hierarchical modelling approach involving all of three modelling methods is briey mentioned, which gives an overview of the design for implementing a generic diagnostic tool on SOFC system. Copyright 2011, Hydrogen Energy Publications, LLC. Published by Elsevier Ltd. All rights reserved.

1.

Introduction

Fuel cell systems are considered as an alternative to conventional fuel combustion power generation, thanks to their lower emissions and higher efciency. Amongst various types of fuel cell, solid oxide fuel cell (SOFC) at high temperature operation allows systems design that well uses the fuel cell thermal output, which leads to higher system efciency than other fuel cell systems such as comparable proton exchange membrane (PEM) fuel cell systems [1]. Due to the importance of efciency and the need to operate fuel

cells at altitude, the hybrid SOFC/gas turbine cycle is a potentially attractive option for applications of auxiliary power unit ofaircraft [1,2] and vehicle as well as for industrial power supply, in stationary and even non-stationary electricity generation applications [3,4]. Besides, SOFCs possess other advantages, i.e. 1. Due to its high temperature operating condition, internal reforming (IR) can be realised; 2. Its insensitivity to gas contaminants enables utilisation of unconventional fuels such as biomass or coal gas;

* Corresponding author. E-mail address: kun.wang@utbm.fr (K. Wang). 0360-3199/$ e see front matter Copyright 2011, Hydrogen Energy Publications, LLC. Published by Elsevier Ltd. All rights reserved. doi:10.1016/j.ijhydene.2011.03.051

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 6 ( 2 0 1 1 ) 7 2 1 2 e7 2 2 8

7213

Additionally, developers expect commercial SOFCs to have lifetimes of 10e20 years, two to four times longer than other fuel cells [5]. However, the capabilities of IR and gas insensitivity lead to more complex electrochemical reactions inside SOFCs. Moreover, carbon formation thermodynamically can take place on the anode material in hydrocarbon-fuelled case. These disadvantages can make SOFC system suffer from a low reliability. In order to avoid catastrophic system failures, an online diagnostic tool for assessing and tracking the state of health of SOFC stack or/and system is very necessary. Nowadays, the diagnosis technique mainly relies on modelbased method [6], by analysing the residuals/deviations of the measured system response from the simulated one by model [7]. Following this diagnostic concept, a reliable and general model which is capable of predicting the normal performance of SOFC is required. In the past decades, a great number of researchers had investigated in SOFC modelling and the internal process simulation based on physical principles. By using physical and analytical equations, they translated successfully the electrochemical reactions, the electronic and ionic properties of materials as well as gas ow process to detailed physical models. These models range from zero-dimensional (0-D) to three-dimensional (3-D) with different features and point to different research objectives. From the viewpoint of model function, 2-D and 3-D modelling is typically concerned with the cell and stack design issues while 0-D and 1-D modelling is aimed at control purposes (on system-level) such as prediction of both the transient and steady-state performance of fuel cell/stack and establishing the optimal operating conditions [5]. For the research target of setting up an online diagnostic tool, low dimensional models (0- and 1-D) are more appropriate due to the less computational time in comparison with the high dimensional ones (2- and 3-D). Moreover, high dimensional models require information about material properties or electrochemical parameters that are not always available or might be difcult to determine. Even so, high dimensional models are still helpful to learn the operation behaviour of fuel cells of different geometry design and very useful for creating training data for black-box modelling which will be introduced in the fth section. Another method is AC impendence modelling. It is based on electrochemical impedance spectroscopy (EIS) measurements. The electrochemical information on an operating fuel cell system can be obtained from the measured EIS data and interpreted by tting this data to an impedance model. Recently, specic applications of EIS in SOFCs have appeared frequently in the literature. The obtained results demonstrate that this technique is an effective modelling approach. It is worth noting that EIS is a tool used to acquire electrochemical parameters. It is also known as AC impedance technique. When a perturbation signal (voltage or current) is imposed on a SOFC, a corresponding output signal (current or voltage) can be obtained. This signal is the reaction of the SOFC to the perturbation. Comparing these two signals can give a characteristic impedance Z(u). In EIS measurement, a series of Z(u) in various frequencies are collected. They are supposed to exhibit the SOFC characteristics and should give information on physical behaviours inside the operating fuel cell.

In fact, both physical and equivalent circuit fuel cell models are mainly based on the knowledge of physicochemical characteristics (electrically, chemically and kinematically), thus also called as white models. They presents a high generalisability level that enables modelling SOFC stacks of different geometric features, but require a high computational effort. In contrast, there is another approach only based on experimental database (no requirement for any physical property), known as the black-box modelling. Black-box models are developed particularly for control-oriented applications, i.e. system monitoring, online control and diagnosis. This approach is appropriate for complex fuel cell system. Nevertheless, the high dependency upon experimental data makes it less generalisable and the fourth approach is thus developed. It falls in between white and black-box approaches, named grey-box modelling. Models based on this method are partially physical and partially empirical. In the following 4 sections (from the 2nd to the 5th section), four modelling approaches for SOFC will be introduced by presenting the models available in literature. It is worth noting that all models reviewed in this paper are with the aim of proposing a state-of-the-art of existing models which may be useful for model-based SOFC system diagnosis. In addition, whichever modelling approach to be used, it should be kept in mind that since phenomena occurring in nature are too complex to be completely described by mathematical equations, the required details to be described by the model must be goal-driven, i.e. the complexity of the model, and the related results, must be strictly connected to the main goal of the analysis itself [8]. In the 6th section, the functions of the reviewed models have been summed up and their application on SOFC stack and system diagnosis is proposed. The validation of models is discussed, too.

2.

Physical models

A great number of papers can be found on SOFC physical modelling. Some were aimed at cell design modication or material development. In this case, the models involve simulations for the temperature distribution, the heat generation, and the ow diffusion. Others focus on predicting cell performance which is expressed either in term of output current density at xed potential or in term of potential at given applied current. Research objective determines the complexity and the dimension of model [9,10]. In this paper, the focus is put on general models which depict cell performance for system analysis. The physical models covered in this section are classied into 4 categories on the basis of model dimensionality and will be introduced in the order of dimension decreasing from 3-D to 0-D. Multi-dimensional (MD) models are set up in the consideration of spatial variation in the physical and chemical variables such as gas concentration, temperature, pressure and current density, for example [8]. In this review, the covered MD models have an identical assumption that the stack is made of repeating single cells stacked together thus a single cell is simulated and its outcome is multiplied by the number of cells to obtain stack results. Such a cell model usually consists of three sub-models, i.e. thermal model, uid

7214

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 6 ( 2 0 1 1 ) 7 2 1 2 e7 2 2 8

model and electrochemical model. The former two (combined together as a thermo-uid model in some literature) are used for the evaluation of temperature prole and molar ux of one or more dimensions inside the fuel cell: the rst calculates temperatures of each component (electrodes and electrolyte); the second calculates gas ow rates as well as their partial pressures at each electrode. The third is an electrical model for predicting cell voltage. It is based on the following relation at a given current density: Ucell Uocv iR hact hconc (1)

where Ucell is the cell voltage, Uocv is the open circuit voltage (OCV), iR is the ohmic drop (or the ohmic polarisation) and hx are the activation and concentration polarisations. The three polarisations contribute to irreversible losses in an operating fuel cell. The OCV equals theoretically to the Nernst potential, related with the local temperature and gas partial pressures. However, in practise, it deviates from this ideal value during cell operation due to concentration drop of reactants. For an overall cell reaction, the cell potential increases with an increase in the activity (concentration) of reactants and a decrease in the activity of products [11]. Zero-dimensional (ZD) models are often used when fuel cell is regarded as a single component of a bigger system, for capturing the general operating behaviour/performance of fuel cell and meeting the requirement of fast computation. When the main purpose of this type of model is to analyse the whole system, the physical-chemical variables variations are not relevant, however, the performances, in terms of power, heat and input requirements are important [12]. Therefore, spatial variation of the parameters considered in MD models can be not taken into account in ZD ones.

2.1.

3-D models

In a SOFC system, the fuel utilisation and the average cell temperature can be controlled by the delivery rate and the temperature of the gases into the cell. If the fuel concentration is high at a cell area, the local electrochemical reaction is active, leading to increased local temperatures and thereby yielding faster reaction rate; in reverse, for a case of fuel depletion, the reaction is inactive, thus a decreased local temperature and a slower reaction rate [13]. Although increased fuel ow tends to increase uniformity of the reaction rates across the active area, it decreases fuel utilisation [13]. Therefore, management of the ow and the inlet temperature of gases is critical to stable cell operation. A 3-D model allows simulating fuel cell internal behaviours and giving information about the temperature and the fuel distributions on three physical dimensions. The nite-volume method is usually employed to separate a unit cell into several parts and thus to simplify the calculation. Ferguson et al. (1996) [14] presented a 3-D SOFC model which could predict the voltage, the mass and electrical distribution at cell-level. The heat source consisted of two terms, i.e. the ohmic heat and the heat from shift and reforming reactants. Since discontinuities of the potential and heat ux at the electrode/electrolyte interfaces exist due to the surface electrochemical reactions, the heat and mass transfer between gas

channels and solid parts were taken into account. A potential drop at the electrode/electrolyte interfaces due to electrochemical reactions was considered. The potential at solid parts was equal to Nernst potential. Faradays law depicts the correlation between the electric current and the mass ux at these interfaces. This model could be used as a design tool to analyse the cell efciency in different geometries (tubular and planar) and congurations (co-, counter-ow and cross-ow designs for a planar geometry). According to the analysis of Ferguson et al., the counter-ow design was considered to be the most efcient for planer geometry which showed less ohmic loss in comparisons with the tubular geometry. However, the effect of radiation was not considered in the model. Yakabe et al. (2001) [15] took into account the radiation mechanism which was regarded as an essential effect on the heat exchange inside the channels while the stack operating at 900e1000  C. At such high temperatures, the excess thermal stress would lead to the non-homogeneous temperature distributions and the mismatch on cell components due to the different thermal expansion coefcients. Therefore, the conducted model in [15] was used to estimate the thermal stresses in the cell components, so as to optimise operating conditions to decrease the temperature gradients of PEN (positive-electrolyte-negative). Paying attention on electrical performance simulation, in the electrochemical model, the concentration polarisation was included in the Nernst potential; the activation polarisation was divided into a constant term and a current-dependent term. The later was integrated into a formula of ohmic resistance. The cell resistance, with the contact resistances included, was estimated from the experimentally measured IeV data of a unit cell. The electric current density was expressed by the Faradays law like in [14]. This model is applicable for SOFCs of counter- and co-ow geometries. Recknagle et al. (2003) [13] proposed a 3-D model for planar SOFC, aiming at investigating the effects of cell ow conguration on the distribution of temperature, current density and fuel distribution. The model can predict the fuel utilisation, the electric current density, and the temperature distribution. In the thermal model, the radiant heat exchange was small and neglected due to the large aspect ratios of the ow channels in the studied stack (length-to-height, roughly 100:1 on the cathode side and 200:1 on the anode side). The electrochemical model could predict local electrical responses to changes in fuel ow rate, local fuel composition, and local temperature. Three geometric congurations were examined on the model and it was concluded that the co-ow case had the most uniform temperature distribution and the smallest thermal gradients. The same investigation was performed by Wang et al. (2007) [16] and the relevant modelling equations can be found in [16].

2.2.

2-D models

The 2-D model is simplied one from the 3-D model by neglecting one dimension. It prompts to some assumptions and simplications, causing a reduction in the resulting information [17]. There are two ways for choosing a 2-D section and they are shown in the following gure (see Fig. 1).

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 6 ( 2 0 1 1 ) 7 2 1 2 e7 2 2 8

7215

Fig. 2 e SOFC equivalent circuit [18].

Fig. 1 e 2-D cross-sections representation of an SOFC unit.

The rst case takes the x-z plane, assuming that all parameters (such as the temperature, the gases concentrations and pressures, etc) are uniform in y direction. In the second case, the current collectors are separated from PEN by gas channels; thereby, the electrical potential on electrode boundary is not constant. The ohmic resistance variation along the neglected direction needs to be taken into account in the model formulation [17]. Xue et al. (2005) [18] developed a dynamic quasi-2D model for a tubular counter-ow SOFC. It was capable of characterising the transient/time- and spatial-dependent properties of critical state variables. The studied section belonged to the second case and was separated, by applying the control volume (CV) method, into four CVs, i.e. anode channel, cell, cathode channel and thermal insulator. the section. Physical properties, within each CV, were assumed to be uniform, but they changed in different CVs. The mass/species balance equation was used to describe the mass/species conservation in each CV. The momentum effect on the main ow stream caused by electrochemical reaction was neglected. The radiation heat transfer between the gas CVs (anode and cathode channel) and the solid CVs (cell and thermal insulator) was not considered but that between the cell CV and its adjacent thermal insulator CV was taken into account in the thermal model. As considering the case that the external load potential was higher than the cell Nernst potential, the fuel cell in this research was regarded as a combination of a Nernst potential source and a capacitor as shown in Fig. 2. Moreover, the three polarisation resistances were calculated based on the instant conditions. As a result, a dynamic model of a tubular SOFC was implemented and it was useful for studying both the steady-state and the transient cell behaviours. It concluded that this model could be used in system optimisation and dynamic controlling. In [19], a model-based study focused on transient operation was carried out based on a dynamic 2-D model. The possibility for improving the cell/stack performance by proper gas ow conguration was conrmed. The possible degradation effects due to increased thermal stresses were pointed out in the

study. The model was validated against empirical data as well as another 1-D model in [20]. Furthermore, the transient behaviour of this model was compared against that of another two-dimensional dynamic model with tubular cell conguration in [21] and the planar stack conguration was found to adapt faster to the operating condition changes. Chnani (2007) [23,22] also took the second case for modelling a planar SOFC with co-ow channels. He developed the thermal and the uidic sub-models through electrical analogy. This method allowed exhibiting gas ows and thermodynamic behaviours in term of equivalent circuit so that multiple identical models could be connected together to be a stack-level model. For a stack module, therefore, this advantage is obvious: the thermal circuit can describe temperature gradient along cells. Fig. 3 shows the schematic diagram for overall cell modelling. The transient thermal model was used to compute the solid and the gas temperatures. The uidic model calculated the partial pressures of chemical species. The electric (electrochemical) model computed the stack voltage and the polarisations with the parameters from these two sub-models. In thermal behaviour modelling [23], the cell was rstly divided into 7 isothermal volumes (as shown in Fig. 4), including anode interconnect, anode channel, electrolyte/ anode interface, electrolyte, electrolyte/cathode interface, cathode channel and cathode interconnect. Fig. 5 shows the 2-D thermal equivalent circuits for each volume. They were connected by temperature nodes to be a nodal network. In order to capture thermodynamic behaviours inside the cell, both heat generation and heat transfer were considered. The former referred to chemical reactions and ohmic losses; it was represented by thermal source symbols. The later was due to three fundamental heat transfer mechanisms (convection, conduction and thermal radiation) as well as mass transfer (heated gas transportation in channels); these four mechanisms were expressed by temperature drops on the corresponding thermal resistances (Rx_conv, Rx_cond, Rx_ray and Rin/out_x illustrated respectively by blue, green, red and black resistance symbols in Fig. 5). The heat capacities at particular nodes indicated thermal energy stored between and inside volumes. In this model, the heat source of chemical reaction is only located at the interface anode/electrolyte. For the uidic model, an equivalent circuit based on the electric uid analogy (as shown in Fig. 6) was built to depict the uidic behaviour in gas channels. The gas ow was homologous to the electrical current while the pressure to the

7216

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 6 ( 2 0 1 1 ) 7 2 1 2 e7 2 2 8

Fig. 3 e Schematic of stack modelling [22].

voltage. The uidic resistance was regarded as the electrical resistance. It was considered that the pressure drop between the air and the fuel sides was linear with the gas ow rate. In the electric model, three modes of polarisation were considered but the resistance of contact was ignored due to its little contribution to the ohmic resistance, when compared with the electrolyte resistance. Finally, a group of the cell-level model could be easily connected in parallel or in series to obtain a stack-level model.

2.3.

1-D models

In 1-D model, the fuel cell is usually treated as a set of layers including interconnects, air channel, electrodes, electrolyte and fuel channel [8], just like the case shown in Fig. 4 but neglecting physical variations at vertical axis. Both gas composition and ow rate in each channel are assumed to be constant and their mean values are used in the simulation. For

planar SOFC, the dimension is following along the gas channel and the direction is determined by the gas ow. It is necessary to note that the fuel cell of cross-ow design cannot be simulated by 1-D models. For tubular SOFC, the kept dimension is usually the tube axis which coincides with the direction of the fuel and oxidant ow [17]. Magistri et al. (2004) [24] built a one-dimensional model for tubular SOFC, where the cell coordinate x is the axis of the tube and its origin corresponds to the bottom of the cell. The main hypotheses of the single cell model are: 1) the cell is adiabatic, 2) the cell voltage is uniform and all the chemical reactions within the anodic stream are at equilibrium, 3) the electrochemical reaction of H2 is taken into consideration; the electrochemical reaction with CO is neglected. The cell model includes: electrochemical performance, equilibrium of reforming and shifting chemical reactions, mass balances of anodic and cathodic, energy balances of gaseous ows, energy balance of the tube and of the solid PEN structure. In the paper, the 1-D model was

Fig. 4 e Heat transfer and heat sources in 7 volumes [22].

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 6 ( 2 0 1 1 ) 7 2 1 2 e7 2 2 8

7217

Fig. 5 e Cell-level thermal model [23].

described and the results were compared to the 0-D model simulation proposed in [25]. In both two models, the input data are geometrical characteristics, operating conditions, inlet ow conditions and gas and material properties. The 1-D model represented the cells as a plug-ow reactor, so it integrated along the cell coordinate the values of the operating. The 0-D model considered the tubular cell as a continuous stirred-tank

Fig. 6 e Electrode uidic model [22].

reactor; and the thermodynamic and electrochemical parameters were thus uniform along the cell coordinate. Both the models were integrated through a relaxation method for the evaluation of the cell performance and were included into a whole system model. The model comparison showed how discriminating the simulation accuracy could be in studied cases. Under some operating condition, it was possible that not great difference was there between the average simulation results from the detailed and the simplied SOFC models; at a rst glance, the results seemed reasonable and compatible with the technological limits of the components of the hybrid system. But a deep analysis of the results from the detailed model revealed that the temperature inside the stack was not uniform and, although the average value was acceptable, the maximum values were too high. The comparison of two different SOFC models is very important for investigating how the study of the whole system is affected by the approximations of SOFC models. Indeed, detailed fuel cell models require a long

7218

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 6 ( 2 0 1 1 ) 7 2 1 2 e7 2 2 8

computational time and knowledge about geometrical data, materials and lay out of the fuel cell, which are rarely available. For these reasons, simplied fuel cell models are generally used in HS simulations, but in this way critical aspect such as hot spot temperature of the SOFC cannot be investigated, and the calculation could give results very different from the real performance of the fuel cell and whole system. On the other hand, there are several cases where the results of the two types of simulation coincide, and it is difcult to have information a priori on the range of operating conditions where this occurs. In [26,27], a so-called dynamic behaviour model of an SOFC was developed and veried. The model was capable of solving the I-V-behaviour and the temperature distribution in the gas ow direction inside a cell operating under either co- or counter-ow mode and it was found to be sufciently accurate for rapid system simulation [28]. The model enabled, e.g., designing the gas ow rates accordingly in respect to the maximum drawn current density and, thereby, to prevent overheating of cell. Aguiar et al. (2004) [29] developed a 1-D dynamic anodesupported intermediate temperature planar SOFC with direct internal reforming. This model predicted the SOFC characteristics both in the steady and the transient states. It consisted of mass and energy balances, and an electrochemical model. For the mass balance the molar ux in the gas channels was considered convective in the ow direction. It was assumed that only hydrogen was electrochemically oxidised and that all of CO was converted through the shift reaction, considered to be at equilibrium. In the fuel channel, three reactions are taken into account: 1) methane steam reforming; 2) water gas-shift; 3) and hydrogen electrochemical oxidation. In the air channel, only the reduction reaction of O2 was considered. Faradays law related the ux of reactants and products to the electric current arising from an electrochemical reaction. In the energy balance were included the released heat from electrochemical reactions and ohmic losses; the convective heat transfer between cell components and gas streams; and the in-plane heat conduction through cell components. The thermal uxes were supposed to be conductive and radiate between the PEN and the interconnect components. However, in the gas channels, they were assumed to be convective in the gas ow direction and from the gas channels to the solid parts. In the electrochemical model the OCV was calculated by the Nernst equation and the SOFC stack was considered isopotential. Costamagna et al. (2004) [31] studied an innovative fuel cell concept, the Integrated Planar Solid Oxide Fuel Cell (IP-SOFC) which was substantially a cross between tubular and planar geometries, seeking to borrow thermal compliance properties from the former and low cost component fabrication and short current paths from the latter. In this new concept, several cells of small dimension were deposited over a vertical porous substrate and electrically connected in series, in order to obtain high voltages and low electrical currents. The scope of IP-SOFC modelling was multi-fold: to better understand the physical-chemical phenomena occurring in the electrodes, in the fuel cell and in the stack, to predict the local behaviour of the cell and to identify dangerous effects (for example, hot spots) which might lead to damage. This model could serve as the basis for planning experimental campaigns and provide

a useful tool for optimisation of fuel cell systems, in aspects of both operating conditions and design parameters. The model involved different levels of simulation: electrode, single cell, tube, bundle, stack and block. At the electrode level, the main phenomena taken into account were the electrochemical reaction, the charge conduction and the mass transfer. At the tube and bundle level, the models included the mass balances of the gaseous streams and energy balances of the gaseous streams and of the solid. Jiang et al. (2006) [30] set up a 1-D dynamic model for a tubular SOFC with external reforming. The cell was divided into elements along the ow direction, like shown in Fig. 7. For each element, there were 4 CVs separated along perpendicular axis (see Fig. 8): the fuel, the solid, the reaction air and the preheated air CVs. Several assumptions were made for the thermal model: 1) for every element, the temperature within each CV was uniform; 2) the radiation and the conduction heat transfer were not taken into account; 3) the convection heat transfer was assumed as the only reason for the temperature gradient of gas streams in the ow. The heat generation due to the reactions (shifting, reforming and electrochemical) and the ohmic losses was calculated. The cell voltage at each element was uniform. An equivalent circuit (see Fig. 9) was built to evaluate the inuence of the current path length to the ohmic loss. This model was capable of predicting SOFC characteristics in both the steady and the transient states and showed a good reliability. Results from the model showed that elevated pressure could improve the cell performance whereas higher operating temperature decreased both the Nernst potential and the irreversible losses (ohmic, activation and concentration losses). Zhang et al. (2006) [32] developed a 1-D non-linear, controloriented dynamic model for planar SOFC. Two kinds of tting function, namely the exponent decay function and the exponent associate function were introduced to t the distribution characteristics of the gaseous molar fractions and the temperature along the streamwise direction. The spatial effect was lumped into the dynamic model by tting the three parameters of the used function. These parameters were determined through numerical simulations. Sorrentino (2006) [33] developed a 1-D steady-state model for co-ow planar SOFC. The model was divided into three sub-models: 1) mass balance model; 2) energy balance model; 3) voltage model. The model was based on the control volume (CV) approach, according to which the cell was discredited in CVs in the ow direction and divided into three layers: anode channel, cathode channel and cell (solid layer). The cell was assumed to be isopotential and the pressure drop across the fuel and air channels was neglected. The radiant heat transfer

Fig. 7 e Schematic of gas ows in a tubular SOFC [30].

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 6 ( 2 0 1 1 ) 7 2 1 2 e7 2 2 8

7219

Fig. 8 e Control volume denitions for one element [30].

and the heat conduction in the solid layer were neglected. The stack was assumed to be adiabatic. The heat convection between solid layer and gas streams and the energy transfer due to the reactants and products were considered dominant in the energy balance. The model showed a good accuracy in the simulation of SOFC states and variables. It was adopted to generate SOFC stack data to be used in a hierarchic modelling approach to implement a control-oriented model [34]. Cheddie et al. (2007) [35] upgraded a 0-D real time model to a dynamic 1-D model in order to predict more accurately the temperature and pressure variations along the gas ow direction. The real time capability was maintained by setting up several simplications: the current density distribution was considered uniform and there was no need of calculating the cell current iteratively, therefore resulting in reduction of computational effort. The overpotentials at each node were replaced by the average one across the cell. It was assumed that the voltage immediately responded to changes in current so the transient states were not taken into account. The gas concentration was considered dependent only on partial pressure rather than both pressure and temperature. In thermal model, all heat generations were assumed to occur in

the PEN. The heat conduction was negligible in the uid phase due to the fact that the thermal conductivity is much higher in the solid regions than in the uid phases. The 1-D model with 21 nodes was proven to require 3.8 ms of computational time for each iteration. The model validation showed that the limiting assumptions did not lead to the signicant simulating difference when comparing with a more comprehensive 1-D model without these assumptions. Moreover, the proposed model was capable of predicting more accurately the transport phenomena. Kang et al. (2009) [36] modied a 1-D dynamic model for a planar internal reforming SOFC also by integrating two simplications: 1) the PEN, interconnects and gas channels were integrated together along the perpendicular direction, that is, the SOFC is considered to have only one temperature layer; 2) the current density distribution is considered to be uniform within the SOFC, and the cell voltage is determined by the average gas molar fractions and cell temperature. These two simplications are similar to the assumptions in Cheddies modelling. In fact, by introducing them, the SOFC model was greatly simplied in form. This model contained 100 nodes and its computational time was decreased comparing with the lumped one. Moreover, it showed an improvement with regard to accuracy because it took into account the spatially distributed nature of SOFCs to a certain extent.

2.4.

0-D models

Fig. 9 e Equivalent circuit for the tubular SOFC cross [30].

The 0-D model is the simplest one. No dimension is determined; and thereby spatial variations are not taken into account. The transformations are considered to dene output variables from input ones. 0-D models are simplied based on assumptions and practical information. They can be used for numerical analysis of fuel cells in energy systems such as SOFC/gas turbine hybrid system. In such a system, the single elements, for instance, compressors, heat exchangers, fuel reformer, partial oxidisers, and contaminant removal apparatus are simulated through independent box models [17]. Furthermore, they allow being easily calibrated and modied for new developed materials. Costamagna et al. (2000) [25] described a hybrid system where the SOFC was simulated with the 0-D model approach. The balance equations were written as macroscopic balances, in form of nite equations. Those equations expressed a balance between inlet and outlet ows of mass and energy in each component of the group; under suitable assumptions, they allowed the evaluation of the average values of the physical-chemical variables of each components and the electrochemical performance of the group itself. In Campanaris 0-D SOFC model (2001) [37], the cell voltage was a function of the current density, the operating temperature and pressure as well as the reactants and product composition. Bove et al. (2005) [12] built a macro model in which the ohmic polarisation depended only upon the material properties. The open circuit voltage and the activation polarisation were related to gas concentration while the concentration polarisation was ignored. The mean current density was regarded as an input variable. Three different modes of gas composition, i.e., inlet gases composition, outlet

7220

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 6 ( 2 0 1 1 ) 7 2 1 2 e7 2 2 8

gases composition and a mean value between the previous two, were respectively used in the simulation for estimating the cell voltage. The simulated results showed that if the rst mode was considered, the effect of fuel utilisation variation could not be estimated; on the contrary, the cell voltage might be underestimated if with the second mode. Therefore, as a conclusion, the third mode seemed as the best, which could get a compromise between these two aspects. Magistri, Ferrari et al. [38e40] made a transient analysis of hybrid system based on SOFC. This system was mainly composed of three parts: the stack, the anodic recirculation system with fuel feeding and the cathodic side (air side) where turbo-machinery and heat exchangers were installed. These transient researches allowed a deep investigation of the Fuel Cell Stack complete with reformer and post-combustor models. In this way it was possible to dene the start-up and shut-down procedures, avoiding risks for the whole plant. The tubular SOFC was considered as a stand-alone unit in order to understand its behaviour without taking into account the inuence of the other components of the plant. This model was developed in the MATLAB-Simulink environment with the TRANSEO [41] tool and was successfully veried at design and off-design conditions. Moreover a special time characterisation of the transient phenomena was introduced in order to automatically suggest to the user the proper integration time step (Dt) to employ. Analysing the results of the model, they observed, during transients condition, unexpected uctuations could occur and need to be carefully monitored, in order to avoid the system running into forbidden or dangerous areas. Modelling the dynamic behaviour of SOFCs is nowadays a highly strategic research area to well address safe operation as well as degradation prevention of SOFC stacks. Bhattacharyya and Rengaswamy, in their extended literature [3], reviewed SOFC dynamic models, highlighted the importance of simulating SOFCs in transient conditions in order to optimise design, control and diagnosis of SOFC systems. With particular regard to these latter goals, on eld of performance monitoring, as well as the management of energy and mass ows during system start-up and load changes, it is required that the development of modelling tools meeting the compromise between satisfactory accuracy and affordable computational burden. The above compromise can be easily achieved by 0-D (i.e. lumped) modelling approaches, thus explaining the increasing number of such scientic contributions that recently appeared in the SOFC literature [3]. An Interesting lumped approach was followed by Sedghisigarchi and Feliachi [42] for control and stability enhancement of SOFC-based distributed generators [43]. Nevertheless, in [42] average cell temperature was assumed as state variable, thus not allowing to provide some basic information for balance of plant analysis, such as temperature of exhaust gases (i.e. outlet SOFC temperature). Sorrentino et al. [34] proposed hierarchical modelling approach to derive a lumped, control-oriented model of planar SOFC capable of accurately simulating temperature and voltage dynamics as function of the main operating variables (i.e. current density, fuel and air utilisations, inlet and outlet temperatures). The contribution [34], differently than [42], does take into account temperature variation across the channels,

thus being suitable to perform, at low computational cost, accurate balance of plant analyses, including heat exchangers sizing [3]. Thus, Sorrentino and Pianese [44] proposed to extend the lumped approach presented in [34] to the modelling of a fully integrated SOFC-APU (i.e. auxiliary power unit). This latter contribution was also proven to be valid for the development of model-based diagnostics tools for mobile SOFCAPUs [45]. The 0-D approach was also applied to transient modelling of tubular SOFC by Hajimoliana [46], to develop suited strategies aimed at controlling voltage and cell-tube temperature by properly acting on both temperature and pressure of the inlet air ow.

3. Equivalent circuit models with EIS technique


EIS is a powerful technique which is usually used to assess the data on the internal resistance, degradation and failure within an electrochemical system. The approach of EIS relies on measuring the impedance of a system at different frequencies by superimposing a small voltage/current AC perturbation onto the voltage/current operating point. The measured impedances are a function of frequency. This data is then used for impedance modelling, including model structure and parameter identication. For fuel cell systems, the measured AC impedances Z(u) are usually displayed through the impedance plot method such as Nyquist plot or Bode plot and three-dimensional perspective plot like in [47]. Impedance modelling is a critical procedure because it is the basis of data interpretation. After data tting, the parameters in the model will be identied. In some cases, the parameters can intuitively provide the information about the processes inside the system; in other cases, however, an interpreting action should be carried out. This phase has a function equal to that of mapping, i.e. transferring from the identied model parameters to the investigated characteristics of the system [48]. In SOFC performance diagnosis, equivalent circuit models are often constructed with electrical elements such as a parallel ReC circuit in the simplest case [49]. Each discrete element is supposed to have corresponding physical meanings and thereby, the interpreting action is not necessary. Once the values of these elements (parameters) are identied in various operating condition, they can be used as an estimated data set for tting physical models. This modelling method is explicated in detail in [50]. Takano et al. (2004) [51] investigated specially the impedance corresponding to mass transfer, usually called gasdiffusion impedance (GDI). In their work, it was assumed that the solid electrolyte, the electrodes, and the current collectors were all of homogeneous composition with uniform thickness. The pressure and the temperature inside the cell were uniform. The ECM with GDI shown in Fig. 10 consists of the resistance (Rwg) and capacitance (Cwg) for GDI, the reaction resistance (Rr), the double layer capacitance (Cd), the electronic resistance (Re), the ionic resistance (Ri) and the ohmic resistance (Rs). The simulation results showed that the GDI was signicantly dependent on the fuel utilisation. The GDI became larger as the fuel utilisation approached both 0 and

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 6 ( 2 0 1 1 ) 7 2 1 2 e7 2 2 8

7221

a grey-box model allows us to extract rules that describe the behaviour of a device. Sorrentino and Pianese [45] presented a grey-box model of a SOFC unit. The core part of the model is the fuel cell stack, made of planar co-ow SOFCs and surrounded by a number of auxiliary devices, namely air compressor/blower, regulating pressure valves, heat exchangers, pre-reformer and postburner. As a consequence of low thermal dynamics characterising SOFCs, a lumped-capacity model is proposed to describe the response of fuel cell and heat exchangers to load change.

5.
Fig. 10 e ECM with GDI [51].

Black-box models

100% and had the inversive relationship with the gas ow rates. Lang et al. (2008) [52] used EIS to measure the different kinds of resistances which contribute to the area specic resistance (ASR) value (or the slope the I-V curve), due to the fact that it is not possible to distinguish them via the IeV curves. Their essential idea is varying the gas composition so that the different processes can be ordered to the different frequency ranges. The measured resistances were tted into the equivalent circuit (see Fig. 11) so as to simulate the electrochemical behaviours of the cells. Among, the impedance of the current collector wires is taken into account by the inductive element ZL; the three terms combining a resistance and a constant phase element (CPE) represent the polarisation of electrodes and the gas conversion impedance; the ohmic resistance is connected in series with these terms.

4.

Grey-box models

The aim of the grey-box modelling is to develop model-based tools aiming at optimal design, management and diagnosis of SOFC units destined to a wide application area [45]. The grey-box method is based on a priori knowledge concerning the process and on the mathematical relations which describe the behaviour of the system. This means that the starting point is a specic model structure based on physical relations. The construction procedure of a grey-box model based on mathematical relations can be divided into different sub-procedures: basic modelling, conduct experiment of the process, calibration and validation [53]. The exibility of

The black-box model is a behavioural model that is derived through statistical data-driven approach. Contrary to the physical models, they are not based on explicit physical equation denitions but the measured database which is capable of reecting the relationship between inputs and outputs. As stated in the previous sections, SOFC is a non-linear, dynamic system with multiple inputs and outputs. So far, most of existing models are based on physical conversion laws and governing equations. Although being useful for analysis and optimisation of SOFC, they are too complex for model-based control system. This drawback impelled some researchers [54e65] to attempt black-box method which is based on mapping inputs to the appropriate outputs. The black-box model is constructed without any physical laws but only a set of input-output pairs for training procedure. It has been veried that the black-box models based articial intelligent are very suitable for non-linear dynamic systems [6]. However, such a model requires an amount of database/ experimental data which should well represent a specic feature of the system. Therefore, the experimental time for collecting data is very long. The following paragraphs will give a summary of this kind of SOFC models most of which aim to predict the cell performance in terms of voltage and/or electric power.

5.1.

ANN (articial neural network)

Fig. 11 e ECM for a SOFC stack [52].

Articial neural network is a statistical data-driven approach. It is inspired by the central nervous system, exploiting features such as high connectivity and parallel information processing, exactly like in the human brain. An ANN is capable of producing a response to a specic combination of input data. With a great number of measured data from a system, the ANN can be trained to learn the internal relationships that govern the system, and then to predict its behaviour at a given input. Across the whole modelling, no relevant physical equation is used. Moreover, the highly parallel connectivity of ANN can reduce the computational time. Arriagada et al. (2002) [54] proposed a non-linear fuel cell model by utilising ANN for evaluating SOFC performance. This model is a two-layer feed-forward network (see Fig. 12). The outputs are air ow, current density, air outlet T (temperature), fuel outlet T, mean solid T, fuel outlet T, mean solid T and the reversible voltage. The model was trained with

7222

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 6 ( 2 0 1 1 ) 7 2 1 2 e7 2 2 8

Fig. 12 e Feed-forward 2-layer neural network.

a reduced amount of input and correct output data pairs generated by a physical cell model. The BP (backpropagation) algorithm was used to modify the weights, which requires the use of differentiable transfer functions (logistic-sigmoid and tanh-sigmoid). Comparing the outputs of the ANN model with that of the physical model, the average values of the errors are well below 1% and the maximum below 4%. Besides the numerical accuracy, the ANN model is much faster and easier to use, which makes it suitable for the generation of performance maps. Milewski et al. (2009) [61] applied an ANN with the same conguration to simulate the SOFC behaviour, using experimental data for training and testing process. This ANN-based SOFC model had 9 input parameters (current density, cathode inlet O2 and N2 ow densities, anode H2 and He ow density, anode thickness, anode porosity, electrolyte thickness and electrolyte temperature) according to which cell voltage could be predicted. A hyperbolic tangent sigmoid transfer function was used as the neuron activation function in the rst layer, whereas a linear transfer function was used in the output layer. The testing results show that ANN can be successfully used in modelling of the singular solid oxide fuel cell. However, its practical design suffers from some drawbacks such as the existence of local minima and over-tting

as well as the determination of the number of hidden layer nodes, etc.

5.2.

RBFNN (radial basis function neural network)

Fig. 13 e RBF neural network.

Compared with the general form of ANN as stated above, radial basis function neural network has a number of advantages, such as better approximation properties, simpler network structures and faster learning algorithms. The RBF neural network is a feed-forward neural network and can uniformly approximate any continuous function to a prespecied accuracy (Warwick, 1996) [63]. It consists of an input layer, a non-linear hidden layer and a linear output layer (see Fig. 13). The input variables are each assigned to nodes in the input layer and connected directly to the hidden layer without weights. The hidden layer nodes (RBF units) calculate the Euclidean distances between the centres and the network input vector, and pass the results through a non-linear function. The output layer nodes are weighted linear combinations of the RBF in hidden layer. The most used non-linear function for RBF units is Gaussian activation function. For realising the RBF algorithm, it is very important that how to choose the optimum initial values of the three parameters: the output weights, the centres and the widths of the Gaussian function. If these parameters are not appropriately chosen, the RBF neural network may degrade validity and accuracy of modelling [64]. The standard training method determines the hidden centres by clustering approach which usually results in a large number of selected centres. In addition, this kind of training method is time-consuming, since it requires examining many different network structures by using a trial and error procedure. Chakraborty built RBF neural networks with 3, 4, 5 and 10 hidden neurons, respectively, in order to nd the optimal network structure for the SOFC modelling [55]. Wu et al. (2007) [64] built a 2-3-1 RBF neural network and utilised a genetic algorithm (GA) to optimise the parameters of the network. The optimum values are regarded as the initial values of the RBFNN parameters and the gradient descent learning algorithms were used to adjust them. GA is a kind of

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 6 ( 2 0 1 1 ) 7 2 1 2 e7 2 2 8

7223

self-adaptive global searching optimisation algorithm based on the mechanics of natural selection and natural genetics [57]. Different from conventional optimisation algorithms, GA is based on population, in which each individual is evolved parallel, and the ultimate result is included in the last population [64]. The similar work was done by Huo et al. (2008) [59] for realising a Hammerstein model of the SOFC in which the nonlinear static part was approximated by an RBFNN and the linear dynamic part was modelled by an autoregressive with exogenous input model. Such a model aimed at controlling fuel utilisation and output voltage of a SOFC stack so that the stack could be protected and the voltage demand of DC type loads could be meted. The natural gas input ow, the oxygen ow, the operating temperature and the stack current were chosen as the model inputs. The fuel utilisation was kept constant. The gas input ow was controlled according to the stack current which was proportional with the load. Through a large number of tests, an RBFNN with 6 hidden nodes was proven for obtaining a better performance. According to Chakraborty [55], the number of hidden layer neurons in RBFNNcan be determined from the training data by a learning algorithm while optimising the number of Gaussian neurons with a global search algorithm, such as genetic algorithm or differential evolution, would cause more computational time to be spent.

1. A rule base containing a number of fuzzy if-then rules; 2. A database which denes the membership functions of the fuzzy sets used in the fuzzy rules; 3. A decision-making unit which performs the inference operations on the rules; 4. A fuzzication interface which transforms the crisp inputs into degrees of match with linguistic values; 5. A defuzzication interface which transform the fuzzy results of the inference into a crisp output [60]. Entchev et al. (2007) [56] applied the Fuzzy Logic Toolbox of MATLAB to build an ANFIS model which could predict SOFC stack current and voltage. This model initialled the parameters in membership function and then they were adjusted by applying a combination of the least squares estimate (LSE) method and the back propagation (BP) gradient decent method. Similarly, Wu et al. (2008) [65] applied ANFIS to build a dynamic model of SOFC stacks for predicting stack voltage. Note that at MATLAB, the ANFIS usually applies a hybrid learning algorithm in which the consequent parameters are identied by the LSE and the antecedent (premise) parameters by the BP.

6.
6.1.

Validation and application of models


Validation of models

5.3.

LS-SVM (least squares support vector machine)

LS-SVM was proposed by Suykens and Vandewalle [62] as a modication of the standard SVM. It possesses prominent advantages over ANN, such as few occurrence of over-tting through the structural risk minimisation principle, and the capability to get the global optimal solution by solving a set of linear equations [65]. A non-linear model of SOFC was established in [58], based on LS-SVM. Fuel utilisation and cell current were chosen as the two inputs and cell voltage as the output. The training data was generated by a mathematical cell model operating at steady-state regime. The RBF function was used as the kernel function of LS-SVM, in which the two important parameters, regularisation parameter and kernel width, were tuned rapidly with a 10-fold cross-validation procedure and a grid search mechanism by LS-SVM toolbox. In comparison with the RBFNN approach, simulation results in this research showed that the LS-SVM yielded higher prediction accuracy.

Model validation is the last step of model development that should always be carried out only after the modelled physical set up is xed. It is usually done by comparing the simulated results with the actual measurement data which were never presented in model parameterisation and tuning. What results to be selected for comparing is dependent on modelling objectives. For system simulation and optimisation, the following parameters may be chosen to be veried: 1. the polarisation curves (IV curves) or/and the impedance spectrum (only for AC impedance modelling approach) where steady-state operating point (such as gas ow rates and compositions, operating temperature, fuel inlet and outlet temperatures) are required; 2. fuel cell performance (such as electrical power or efciency, operating voltage and temperature prole) at different load and operating conditions; 3. transient behaviour of the stack (stack temperature evolution against time, voltage or current prole) during load changes. If a model can reproduce the same results (at a given level of condence) with the experimental ones, this model is considered valid. However, some degree of uncertainty in experiments and also in calculations always exists [8] which might result in large deviation between the computed and the measured values. In practise the data obtained from even repeated experiments are never identical while in simulation the model should always produce the same output when given the same input. Due to this fact, the uncertainty of measurement data should be considered when validating a model. In [18], the

5.4.

ANFIS (adaptive neural-fuzzy inference system)

ANFIS is a fuzzy inference system (FIS) implemented in the framework of adaptive networks. It was put forward by Dr. Jang while various combinations of methodologies in soft computing emerged. It integrates the advantage of both neural networks and fuzzy system, which not only has good learning capability but can also be interpreted easily [65]. The architecture of ANFIS and the methods to update parameters in membership functions during learning process have been introduced in detail in [60]. The FIS is composed of ve functional blocks (see Fig. 14):

7224

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 6 ( 2 0 1 1 ) 7 2 1 2 e7 2 2 8

conducted model was validated at steady-state condition by verifying the correlation of the experimental polarisation and power density curves with the simulated ones. The model was nally considered valid because the simulation results had a consistent trend when compared to experimental data. Furthermore, due to high temperature operating, there is a large chance of having some variation in materials resulting in defects in the SOFC structure. Hence, a lower experimental performance must be expected. In [31], the authors considered during model validation the effects of micro-cracks which lead to 1) a not perfect adherence between cell components and 2) some cross-over of the reactant gases between two electrode sides. These effects were simulated respectively 1) by introducing an additional constant resistance into the model and 2) by considering occurrence of a chemical reaction between oxygen and hydrogen which causes the presence of water in both the anodic and the cathodic ow rates. The polarisation curves obtained from the simulations and the experiments are compared to verify the agreement of these results. Besides the uncertainty of experiment, another difculty of SOFC model validation is due to limited resources and measurement techniques such as the difculties of measuring variables like local current density, temperature or gas composition, especially for multi-dimensional SOFC models. In this case, an indirect validation can be performed by comparing the predictions of two or more independent models for an identical test case [8]. Achenbach (1994) [66] compared eight independent models for a predened benchmark test. The values of three outputs from these models, i.e. the maximum solid temperature, minimum solid temperature and the air exit temperature, were analysed by statistical method. The 2-D model in [19] was validated against a 1-D model as well as empirical data. The validation of the 2-D model in [67] was carried out with measurements and a 3-D computational uid dynamics model. The 1-D model in [24] was validated by being compared with a 0-D model. In [12], the authors compared three 0-D models with different gas composition modes as an input for simulating SOFC performance evaluation. Through analysing the results with physical and empirical knowledge on SOFC system, it was concluded that using the average composition rather than inlet or outlet gases composition when the fuel utilisation is low, the SOFC behaviour could be approximated well. For black-box models, before modelling, experimental data are divided into three parts the third of which is for

model validation. It should be ensured on one hand that none of points in such a data set is involved in the training data set and on the other hand that this validation data set should be involved in the mathematical space of the training one. Therefore, validation of black-box model permits to assess how well the model can explain the signicant information in the training data and how it will generalise to an independent data set with the same or similar information involved in.

6.2.

Application of models for SOFC system diagnosis

knowledge base database input rule base output

fuzzification interface

defuzzification interface

fuzzy

decision-making unit

fuzzy

Fig. 14 e Fuzzy inference system [60].

As stated in the introduction, this review is aimed to propose a state-of-the-art of existing SOFC models for system modelbased diagnosis. The design for a fault-diagnosis system begins frequently with mathematical modelling of the process [68], following the idea of comparing the measured response of system with the simulated normal one to determine whether or not a fault has occurred during the real process. In our case, the diagnosed object will be the FC system with potential failures in operation that yield a sudden drop of performance. The stack will be treated as a sensor. The output deviation from the model will be analysed in order to assess system performance. Setting up such a generic diagnostic tool requires an intensive model use for example by combining fast models with a statistical representation of both operative and state variables, and in this case black-box and grey-box models are more appropriate than physical ones [45]. Black-box model is very suited for interpreting and predicting the performance of SOFC at both normal and abnormal operating conditions and of different geometric designs, avoiding using complicated differential equations to describe the stack [64]. As long as the input-output data for depicting the studied process is available, an accurate model of this type can be achieved relying upon optimisation on training algorithm or/and model architecture. Unfortunately, the availability of experimental data that should be enough representative for the modelled phenomena is always a major problem of black-box modelling, especially in SOFC modelling for diagnostic application due to the high risk of fuel cell damage and experimental failure when a system fault occurs (for example a low ow rate of input or a too high operating temperature which may be catastrophic for fuel cell and lead to system shut-down). Overcoming this problem needs to borrow validate white-box models to supplement the experimental data matrix. In [54,55,58,59,64,65], the black-box models were trained and validated by using pure data produced from physical models. These models showed a good congruence with the physical ones and to be generic to various operational conditions as well. The performances of them are, however, limited by the accuracy of the applied physical models and should be thus further validated and modied with experimental data. It is worth noting that all the reviewed black-box models in this paper are straightforward type aiming at simulating static SOFC processes. However, the application of neural networks to the modelling or fault diagnosis of control systems requires taking into account the dynamics of processes or system considered [6]. Hence, SOFC recurrent neural network models that include feedback loops from output would be very

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 6 ( 2 0 1 1 ) 7 2 1 2 e7 2 2 8

7225

Fig. 15 e Interactions and Applications of different model types in online SOFC system diagnosis.

necessary to be considered and developed. In fact, this type of model had been developed for proton exchange membrane (PEM) fuel cell dynamic modelling by Jemei et al. (2004) [69] and Puranik et al. (2010) [70] but at present, there is no such model for SOFC dynamics in the available literature. Accordingly, recurrent neural network for SOFC dynamic modelling will carry more weight in the future work. Different from black-box model, grey-box model requires both the knowledge on the process or studied system (represented by mathematical relations) and empirical data. Moreover, the more the model is detailed, the more data/ information is required, especially with respect to geometric design and material properties. However, a reliable grey-box model for fuel cell can be realised also with a general knowledge of the geometry and an empirical denition of the losses as a function of stack operating temperature, e.g. area specic resistance. The existing dynamic 2-D and 1-D models can be used as a starting point in the development of a grey-box model. For diagnosis application, it can be assumed in such a model that SOFC behaves as a rst-order system and thermal dynamics is much slower and thus dominant with respect to the dynamics of electrochemistry and mass transfer. In this case, SOFC is simulated by applying the conservation of energy principle (heat balance) for a lumped control volume, which includes air and fuel channels, as well as interconnect and solid tri-layer (i.e. electrolyte and electrodes) [33,44,34].

7.

Conclusion & prospective

In this paper four modelling approaches for SOFC are presented. The physical models have been greatly developed in recent decade. They are constructed based on the physical laws in thermo-uid and electrochemical courses, which can describe the internal physical and chemical behaviours inside

a fuel cell unit. The 3-D and 2-D models are complicated and time-consuming in spite of their considerable accuracy. The improved 1-D and 0-D models by being integrated several simplications are regarded as the most pertinent for real time simulation applications. These simplications (or assumptions) are required to not only reduce the computational time but improve the model reliability as well. They are suitable for those applications requiring a satisfactory compromise between accuracy and computational time, such as model-based control and diagnostics. The equivalent circuit models can be used to simulate the AC impedance spectrum of SOFCs, providing information about the individual behaviours inside a cell or a stack. However, due to the complexity and coupling of physicochemical processes in SOFCs, it remains technical gaps in the AC impedance modelling and simulation in terms of fundamental understanding [48]. The grey-box method is based on a combination of a priori knowledge concerning the process and the mathematical relations which describe the behaviour of a SOFC system. Its exibility allows us to extract rules that describe or interpret the behaviours. The black-box models can predict the fuel cell performance without knowledge of numerous physical, chemical and electrochemical parameters. The underlying self-learning process ensures adapting the model to new situations. The black-box models based on articial intelligence are exible and pertinent for the non-linear dynamic FC system [6]. The disadvantages are that 1) the collection of experimental database should be perform in a long time and 2) a raw data processing is necessary in order to reduce the training time and improve the delity of the model. The validation of models of each type has been discussed. With the aim of setting up a generic model-based diagnostic tool for SOFC systems, the roles of models within the diagnostic algorithm development have been pointed out. Dynamical simulation of system process is required for

7226

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 6 ( 2 0 1 1 ) 7 2 1 2 e7 2 2 8

realising online fault diagnosis, which may lead to selecting recurrent neural network for SOFC modelling. Considering the limited information level retained by the experimental data, it will be taken into account the extension of the available data set by means of hierarchical modelling approach in future work. An overall prospective for real time model-based diagnosis of SOFC system is sketched in Fig. 15. The black-box model is expected for static and especially dynamic simulation of SOFC responses to inputs. The grey-box model developed from 1-D or 2-D physical models is aimed to describe thermal dynamics of fuel cell and to represent the balance of plant, providing details in physical sense. 3-D and 0-D models are two extremes in physical models and they will be used to supplement experimental data matrix (set up based on design of experiment (DOE)) for reinforcing the black-box model applicability. The former aims at various geometries of SOFCs and the later at different operating conditions. In addition, these data may also be useful for knowledge/feature extraction of considered faults, serving for fault identication and localisation (refer to [68]) in the last diagnosis step. The equivalent circuit models based on EIS technique will be used for stack degradation analysis in frequency domain to know in what frequency range a given fault can be observed. These results can be used for signal-based diagnosis (refer to [68]).

Acknowledgements
The nancial support of the European Commission for the GENIUS Collaborative Project is gratefully acknowledged.

references

[1] Freeh JE, Pratt JW, Brouwer J. Development of a solid-oxide fuel cell/gas turbine hybrid system model for aerospace applications, In: proceedings of ASME Turbo Expo, Power for Land, Sea, Air 2004, June 14e17, Vienna, Austria. [2] Steffen Jr CJ, Freeh JE, Larosiliere LM. Solid oxide fuel Cell/Gas turbine hybrid cycle Technology for auxiliary Aerospace power. Reno, Nevada: Prepared for AMSE Turbo Expo; 2005. June 6-9, 2005. [3] Bhattacharyya D, Rengaswamy R. A review of solid oxide fuel cell (SOFC) dynamic models. Industrial & Engineering Chemistry Research 2009;48:19. [4] Yamada T, Chitose N, Akikusa J, Murakami N, Akbay T, Miyazawa T, et al. Development of intermediatetemperature SOFC module using doped lanthanum gallate. Pennington, NJ: Electrochemical Society; 2003. ETATS-UNIS, pp. XX, 1520 pp. [5] Braun RJ. Optimal design and operation of solid oxide fuel cell systems for small-scale stationary applications, Doctoral Dissertation. University of Wisconsin-Madison; 2002. [6] Patan K. Articial neural networks for the modelling and fault diagnosis of technical process, Lecture notes in control and information Sciences, vol. 377. Springer; 2008. p. 206. [7] Chen J, Patton RJ. Robust model-based fault diagnosis for dynamic systems. Norwell, MA, USA: Kluwer Academic Publishers; 1999. . [8] Bove R, Ubertini S. Modeling solid oxide fuel cells: methods, procedures and techniques. In: Bansal NP, editor. Fuel cells and Hydrogen energy. Springer; 2008.

[9] Murshed AM, Huang B, Nandakumar K. Control relevant modeling of planer solid oxide fuel cell system. Journal of Power Sources 2007;163:830e45. [10] Petruzzi L, Cocchi S, Fineschi F. A global thermoelectrochemical model for SOFC systems design and engineering. Journal of Power Sources 2003;118: 96e107. [11] Eg Services G. Fuel cell handbook [electronic resource] / EG&G Technical Services, Inc. Morgantown, WV: U.S. Dept. of Energy, Ofce of Fossil Energy, National Energy Technology Laboratory; 2004. [12] Bove R, Lunghi P, Sammes MN. SOFC mathematic model for systems simulations. Part one: from a micro-detailed to macro-black-box model. International Journal of Hydrogen Energy 2005;30:181e7. [13] Recknagle KP, Williford RE, Chick LA, Rector DR, Khaleel MA. Three-dimensional thermo-uid electrochemical modeling of planar SOFC stacks. Journal of Power Sources 2003;113: 109e14. [14] Ferguson JR, Fiard JM, Herbin R. Three-dimensional numerical simulation for various geometries of solid oxide fuel cells. Journal of Power Sources 1996;58:109e22. [15] Yakabe H, Ogiwara T, Hishinuma M, Yasuda I. 3-D model calculation for planar SOFC. Journal of Power Sources 2001; 102:144e54. [16] Wang G, Yang Y, Zhang H, Xia W. 3-D model of thermo-uid and electrochemical for planar SOFC. Journal of Power Sources 2007;167:398e405. [17] Bove R, Ubertini S. Modeling solid oxide fuel cell operation: approaches, techniques and results. Journal of Power Sources 2006;159:543e59. [18] Xue X, Tang J, Sammes N, Du Y. Dynamic modeling of single tubular SOFC combining heat/mass transfer and electrochemical reaction effects. Journal of Power Sources 2005;142:211e22. [19] Kemm M, Selimovic A, Assadi M. A dynamic model of an Atmospheric solid oxide fuel cell system for stationary power generation, 4th International ASME Conference on fuel cell Science, engineering and Technology. Irvine, CA, United States: ASME Press; 2006. [20] Kemm M. Dynamic solid oxide fuel cell modelling for nonsteady State simulation of system applications, Faculty of engineering LTH Department of energy Sciences Division of thermal power engineering; 2006. [21] Kemm M, Stiller C, Selimovic A, Thorud B, Torisson T, Bolland O. Planar and tubular solid oxide fuel cells: a comparison of transient process behaviors. Pennington, NJ: Electrochemical Society; 2005 [ETATS-UNIS]. ra M-C, Glises R, Kauffmann JM. Transient [22] Chnani M, Pe thermal behaviour of a solid oxide fuel cell, the 5th International Conference on fuel cell Science. Engineering and Technology. Fuel Cell; 2007. 2007, [New York]. lisation macroscopique de piles PEFC et [23] Chnani M. Mode tude de leur couplage, The ` se de doctorat, SOFC pour le partement ENISYS UMR CNRS 6174; Institut FEMTO-ST-De UFC-UTBM-ENSMM, 2008. [24] Magistri L, Bozzo R, Costamagna P, Massardo AF. Simplied versus detailed solid oxide fuel cell reactor models and inuence on the simulation of the design point performance of hybrid systems. New York: American Society of Mechanical Engineers; 2004 [ETATS-UNIS]. [25] Costamagna P, Magistri L, Massardo AF. Design and part-load performance of a hybrid system based on a solid oxide fuel cell reactor and a micro gas turbine. Journal of Power Sources 2001;96:352e68. [26] Gubner A. Non-isothermal and dynamic SOFC voltagecurrent behavior. Pennington, NJ: Electrochemical Society; 2005 [ETATS-UNIS].

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 6 ( 2 0 1 1 ) 7 2 1 2 e7 2 2 8

7227

[27] Gubner A, Froning D, Haart BD, Stolten D. Complete modeling of kW-range SOFC stacks. Pennington, NJ: Electrochemical Society; 2003 [ETATS-UNIS]. [28] Gubner A, Saarinen J, Ylijoki J, Froning D, Kind A, Halinen M, et al. Dynamic co-simulation of a solid oxide fuel cell (SOFC) and the balance of plant (BoP) by combining an SOFC model with the BoP-modelling tool APROS, 7th European fuel cell Forum; 2006 (EFCF 2006) 2006, Lucerne, Switzerland. [29] Aguiar P, Adjiman CS, Brandon NP. Anode-supported intermediate temperature direct internal reforming solid oxide fuel cell. I: model-based steady-state performance. Journal of Power Sources 2004;138:120e36. [30] Jiang W, Fang R, Khan JA, Dougal RA. Parameter setting and analysis of a dynamic tubular SOFC model. Journal of Power Sources 2006;162:316e26. [31] Costamagna P, Selimovic A, Del Borghi M, Agnew G. Electrochemical model of the integrated planar solid oxide fuel cell (IP-SOFC). Chemical Engineering Journal 2004;102: 61e9. [32] Zhang X, Li J, Li G, Feng Z. Development of a control-oriented model for the solid oxide fuel cell. Journal of Power Sources 2006;160:258e67. [33] Sorrentino M. Development of a hierarchical structure of models for simulation and control of planar solid oxide fuel cells. Italy: Department of Mechanical Engineering, University of Salerno; 2006. [34] Sorrentino M, Pianese C, Guezennec YG. A hierarchical modeling approach to the simulation and control of planar solid oxide fuel cells. Journal of Power Sources 2008;180: 380e92. [35] Cheddie DF, Munroe NDH. A dynamic 1D model of a solid oxide fuel cell for real time simulation. Journal of Power Sources 2007;171:634e43. [36] Kang YW, Li J, Cao GY, Tu HY, Li J, Yang J. A reduced 1D dynamic model of a planar direct internal reforming solid oxide fuel cell for system research. Journal of Power Sources 2009;188:170e6. [37] Campanari S. Thermodynamic model and parametric analysis of a tubular SOFC module. Journal of Power Sources 2001;92:26e34. [38] Ferrari ML, Traverso A, Massardo AF. Transient analysis of solid oxide fuel cell hybrids: Part B e Anode recirculation model. ASME Conference Proceedings 2004; 2004. 399e407. [39] Magistri L, Ferrari ML, Traverso A, Costamagna P, Massardo AF. Transient analysis of solid oxide fuel cell hybrids: Part C e Whole-Cycle model. ASME Conference Proceedings 2004; 2004. 525e532. [40] Magistri L, Trasino F, Costamagna P. Transient analysis of solid oxide fuel cell Hybridse-Part I: fuel cell models. Journal of Engineering for Gas Turbines and Power 2006;128:288e93. [41] Traverso A. TRANSEO Code for the Dynamic Performance Simulation of Micro Gas Turbine Cycles. ASME Conference Proceedings 2005; 2005, 45e54. [42] Sedghisigarchi K, Feliachi A. Dynamic and transient analysis of power distribution systems with fuel Cells-part I: fuel-cell dynamic model. Energy Conversion, IEEE Transactions on 2004;19:423e8. [43] Sedghisigarchi K, Feliachi A. Dynamic and transient analysis of power distribution systems with fuel Cells-part II: control and stability enhancement. Energy Conversion, IEEE Transactions on 2004;19:429e34. [44] Sorrentino M, Pianese C. Control oriented modeling of solid oxide fuel cell auxiliary power unit for transportation applications. Journal of Fuel Cell Science and Technology 2009;6:041011. [45] Sorrentino M, Pianese C. Grey-Box modeling of SOFC unit for design, control and diagnostics applications. Lucerne, Switzerland: European Fuel Cell Forum; 2009. June 29-July 2 2009.

[46] Hajimolana SA, Soroush M. Dynamics and control of a tubular solid-oxide fuel cell. Industrial & Engineering Chemistry Research 2009;48:6112e25. [47] Nahir TM. Impedance spectroscopy: Theory, experiment, and applications. 2nd ed Edited by Evgenij Barsoukov (Texas Instruments Inc.) and J. Ross Macdonald (University of North Carolina, Chapel Hill). John Wiley & Sons, Inc.: Hoboken, NJ. 2005. xvii 596 pp. $125.00. ISBN 0471-64749-7. Journal of the American Chemical Society 2005;127:12431. [48] Huang Q-A, Hui R, Wang B, Zhang J. A review of AC impedance modeling and validation in SOFC diagnosis. Electrochimica Acta 2007;52:8144e64. [49] Singhal SC, Kendall K. High-temperature solid oxide fuel cells: Fundamentals, design and applications. 1 edition ed. Elsevier Science; 2003. [50] Leonide A, Apel Y, Ivers-Tiffee E. SOFC modeling and parameter identication by means of impedance spectroscopy. ECS Transactions 2009;19:81e109. [51] Takano K, Nagata S, Nozaki K, Monma A, Kato T, Kaga Y, et al. Numerical simulation of a disk-type SOFC for impedance analysis under power generation. Journal of Power Sources 2004;132:42e51. [52] Lang M, Auer C, Eismann A, Szabo P, Wagner N. Investigation of solid oxide fuel cell short stacks for mobile applications by electrochemical impedance spectroscopy. Electrochimica Acta 2008;53:7509e13. [53] Sohlberg B. Grey box modelling for model predictive control of a heating process. Journal of Process Control 2003;13: 225e38. [54] Arriagada J, Olausson P, Selimovic A. Articial neural network simulator for SOFC performance prediction. Journal of Power Sources 2002;112(7). [55] Chakraborty UK. Static and dynamic modeling of solid oxide fuel cell using genetic programming. Anglais 2009;34:12. [56] Entchev E, Yang L. Application of adaptive neuro-fuzzy inference system techniques and articial neural networks to predict solid oxide fuel cell performance in residential microgeneration installation. Journal of Power Sources 2007; 170:8. [57] Goldberg DE. Genetic algorithms in search, optimization and machine learning. Addison-Wesley Educational Publishers Inc; 1989. [58] Huo HB, Zhu XJ, Cao GY. Nonlinear modeling of a SOFC stack based on a least squares support vector machine. Journal of Power Sources 2006;162:6. [59] Huo HB, Zhu XJ, Hu WQ, Tu HY, Li J, Yang J. Nonlinear model predictive control of SOFC based on a Hammerstein model. Journal of Power Sources 2008;185:7. [60] Jang JSR. ANFIS: adaptive-network-based fuzzy inference system. Systems, Man and Cybernetics, IEEE. Transactions on 1993;23:665e85. [61] Milewski J, Swirski K. Modelling the SOFC behaviours by articial neural network. International Journal of Hydrogen Energy 2009;34:8. [62] Suykens JAK, Vandewalle J. Least squares support vector machine Classiers. Neural Processing Letters 1999;9: 293e300. [63] Warwick K, Craddock R. An introduction to radial basis functions for system identication. A comparison with other neural network methods. Proceedings of the 35th IEEE 1996. In: Decision and Control, vol. 461; 1996. 464e469. [64] Wu XJ, Zhu XJ, Cao GY, Tu HY. Modeling a SOFC stack based on GA-RBF neural networks identication. Journal of Power Sources 2007;167:145e50. [65] Wu XJ, Zhu XJ, Cao GY, Tu HY. Nonlinear modeling of a SOFC stack based on ANFIS identication. Simulation Modelling Practice and Theory 2008;16:399e409. [66] Achenbach E. Status of the IEA-Bench Mark test 1 on stack modelling. Feb. 22-25. Rome: IEA. Workshop; 1994.

7228

i n t e r n a t i o n a l j o u r n a l o f h y d r o g e n e n e r g y 3 6 ( 2 0 1 1 ) 7 2 1 2 e7 2 2 8

[67] Larrain D. Solid oxide fuel cell stack simulation and optimization, including experimental validation and transient behavior, Ph.D Thesis, Institut des science de nergie; Ecole polytechnique fe de rale de lausanne, le Lausanne, 2005. [68] Isermann R. Fault-Diagnosis systems: an introduction from fault Detection to fault Tolerance. Springer; 2005.

[69] Jemei S, Hissel D, Pera MC, Kauffmann JM. Dynamical recurrent neural network towards modeling of on-board fuel cell power supply. Industrial Electronics 2004;vol. 471:471e6. IEEE International Symposium on 2004. [70] Puranik SV, Keyhani A, Khorrami F. Neural network modeling of proton exchange membrane fuel cell. Energy conversion. IEEE Transactions on 2010;25:474e83.