DE102020208540A1 - Method and device for operating a fuel cell system using machine learning methods - Google Patents

Abstract

The invention relates to a computer-implemented method for determining a system state variable (x(i+1)) in a fuel cell system (1) at successive evaluation times, the following steps being carried out at each current evaluation time (i+1): - determining (S2) one or more operating state variables (y(i+1)) of the fuel cell system (1) at the current evaluation time (i+1);- determining (S3) a current system state variable (x(i+1)) using a trained recurrent neural network depending on an internal state vector (h(i)) of the recurrent neural network, which indicates an internal state of the fuel cell system (1) at a previous evaluation time (i), and on the determined operating state variables (y(i+1) ) at the current evaluation time (i+1); and - operating (S3) the fuel cell system (1) depending on the currently determined system state variable (x(i+1)).

Description

technical field

The invention relates to fuel cell systems, in particular SOFC systems (SOFC: Solid Oxide Fuel Cell) for operating a fuel cell system with natural gas. In particular, the present invention relates to determining a system state of a reformer device for processing the hydrocarbons contained in the natural gas into hydrogen-rich gas.

Technical background

Conventional fuel cell systems that operate on natural gas include a reformer device that is used to convert longer chain hydrocarbons contained in natural gas into a hydrogen-rich gas. The hydrogen-rich gas is then oxidized with atmospheric oxygen in the so-called fuel cell stack to generate electrical energy.

In addition to the natural gas, exhaust gas from the cathodes of the fuel cell stack is also recirculated to the reformer device used to convert the hydrocarbons into natural gas. The exhaust gas recirculation rate can be set variably.

Due to the exhaust gas recirculation, the fuel cell system has a highly non-linear behavior that is determined by an internal system state of the fuel cell system. This internal system state depends significantly on the course of the operation of the fuel cell system. The internal system state of the fuel cell system is specified by at least one system state variable, which can include an oxygen-to-carbon ratio at a specific point in the fuel cell system, in particular on the input side of the reformer device, and/or a fuel utilization rate. In particular, the system state variable is the basis of an operating mode, in particular a control or regulation of the fuel cell system.

However, this system state variable cannot be measured online, but can only be determined by extensive measurement of the gas compositions. These measurements are time-consuming and cannot be used for on-the-fly estimations of the internal system state. Physical modeling of the oxygen-to-carbon ratio and the fuel economy rate is possible, but currently not suitable for on-line application due to the complexity and long calculation time, and in particular not suitable for use in closed-loop control. Exact measurements are also not possible in real time.

In current fuel cell systems, the system state variable in the form of the oxygen-to-carbon ratio and the fuel efficiency is determined in order to operate the fuel cell system. The system state variable is determined using an imprecise physical model. Therefore, they only indicate the current system status of the reformer device imprecisely.

Disclosure of Invention

According to the invention, a method for operating a fuel cell system with a reformer device according to claim 1 and a device for operating a fuel cell system and a fuel cell system according to the independent claims are provided.

Further developments are specified in the dependent claims.

• - Ermitteln von einer oder mehrerer Betriebszustandsgrößen des Brennstoffzellensystems zu dem aktuellen Auswertungszeitpunkt;
• - Ermitteln einer aktuellen Systemzustandsgröße mithilfe eines trainierten rekurrenten neuronalen Netzwerks abhängig von einem Interner-Zustands-Vektor des rekurrenten neuronalen Netzwerks, der einen internen Zustand des Brennstoffzellensystems angibt, zu einem vorangehenden Auswertungszeitpunkt, und von den ermittelten Betriebszustandsgrößen zu dem aktuellen Auswertungszeitpunkt; und
• - Betreiben des Brennstoffzellensystems abhängig von der aktuell ermittelten Systemzustandsgröße.

According to a first aspect, a method is provided for determining a system state variable in a fuel cell system at successive evaluation times, the following steps being carried out at each current evaluation time:

• - Determination of one or more operating state variables of the fuel cell system at the current evaluation time;
• - Determining a current system state variable using a trained recurrent neural network depending on an internal state vector of the recurrent neural network, which indicates an internal state of the fuel cell system, at a previous evaluation time, and from the determined operating state variables at the current evaluation time; and
• - Operation of the fuel cell system depending on the currently determined system state variable.

Controlling a fuel cell system with exhaust gas recirculation requires knowledge of an internal system state in the form of a system state variable, such as an oxygen-to-carbon ratio and/or the fuel utilization rate. However, these variables can only be determined based on measurements of the gas compositions of gases supplied by the fuel cell. However, such measurements are complex and cannot be carried out while the fuel cell is in operation. The system state variable, namely the oxygen-to-carbon ratio and the fuel utilization rate, can be modeled physically, but this modeling is too imprecise for a reliable control of the fuel cell system. Exact measurements are possible, but expensive. In particular for controlling or regulating the fuel cell system, it is necessary to know the system state variable.

An idea of the above method is to find the system state variable, namely the oxygen-to-carbon ratio or the fuel utilization rate, using a recurrent neural network. For this purpose, measured variables that are present in the fuel cell system, such as temperature, volumetric mass flows, gas composition of the supplied natural gas and the like, as well as valve manipulated variables, are to be used as operating state variables.

The system state variable can be modeled using a physical model that is known per se. However, because of tolerances in the operating state variables and because of inaccuracies in the physical model, there are deviations in the estimation of the system state variable, which accumulate and possibly lead to considerable inaccuracy in the estimation of the internal system state of the fuel cell system. Therefore, according to the above method, it is proposed to make the physical model more precise using a recurrent neural network.

The combination with the physical model makes it possible to train the recurrent neural network with a smaller number of training data sets. This is particularly advantageous since determining the system state of a fuel cell system is very complex and is only possible for a small number of reference measurements for the purpose of training the recurrent neural network.

The use of the physical model allows the use of knowledge about known physical relationships to determine a system state of the fuel cell system. The system state variable determined in this way can then be made more precise by a corrective intervention of the trained recurrent neural network. The hybrid model formed in this way makes it possible to make good predictions of the system state of the fuel cell system even when training the recurrent neural network with few training data sets. Overall, the above method makes it possible to provide a system state based on the system state variable, which could otherwise only be physically modeled very imprecisely, so that the fuel cell system can be controlled or regulated in an improved manner based on the system state variable.

Furthermore, the determination of the current system state variable can continue to be carried out as a function of a progression of the operating state variables over a number of past evaluation times.

It can be provided that the operating state variables include at least one of the following variables: a generated current, a temperature of a fuel cell unit of the fuel cell system, one or more volume flows and a fuel composition at the inlet of the fuel cell system and one or more manipulated variables, in particular valve manipulated variables for controlling flow rates, include or depend on them.

According to one embodiment, the trained recurrent neural network can be designed to determine a system state variable correction depending on the internal state vector of the recurrent neural network at the previous evaluation time and on the operating state variables at the current evaluation time, with which one with the operating state variables using a physical model modeled system state variable is applied to the current evaluation time to determine the current system state variable.

In particular, the trained recurrent neural network can take into account the system state variable modeled using the physical model at the current evaluation time in order to determine the system state variable correction (K).

Furthermore, the trained recurrent neural network can have a GRU unit or an LSTM unit.

According to one embodiment, the fuel cell system can be operated as a function of the current system state variable, with one or more of the variables a recirculation mass flow, a fuel mass flow and an oxygen mass flow being set variably as manipulated variables depending on the current system state variable.

• - Ermitteln von einer oder mehrerer Betriebszustandsgrößen des Brennstoffzellensystems zu einem aktuellen Auswertungszeitpunkt;
• - Ermitteln einer aktuellen Systemzustandsgröße mithilfe eines trainierten rekurrenten neuronalen Netzwerks abhängig von einem Interner-Zustands-Vektor des rekurrenten neuronalen Netzwerks, der einen internen Zustand des Brennstoffzellensystems zu einem vorangehenden Auswertungszeitpunkt angibt, und von den Betriebszustandsgrößen zu dem aktuellen Auswertungszeitpunkt; und
• - Betreiben des Brennstoffzellensystems abhängig von der aktuell ermittelten Systemzustandsgröße.

According to a further aspect, a device, in particular a control unit, is provided for operating a fuel cell system, with a current system state variable being determined in the fuel cell system, with the device being designed to cyclically carry out the following steps:

• - Determination of one or more operating state variables of the fuel cell system at a current evaluation time;
• - Determining a current system state variable using a trained recurrent neural network depending on an internal state vector of the recurrent neural network, which indicates an internal state of the fuel cell system at a previous evaluation time, and on the operating state variables at the current evaluation time; and
• - Operation of the fuel cell system depending on the currently determined system state variable.

According to a further aspect, a fuel cell system is provided, comprising a fuel cell; a reformer device; an exhaust gas recirculation; a sensor system for measuring measured variables as operating state variables; and the above device.

character list

• 1 eine schematische Darstellung eines Brennstoffzellensystems;
• 2 eine schematische Funktionsdarstellung zur Ermittlung des Systemzustands des Brennstoffzellensystems;
• 3 eine beispielhafte Darstellung einer GRU-Einheit als eine RNN-Einheit; und
• 4 ein Flussdiagramm zur Veranschaulichung des Verfahrens zum Ermitteln der Systemzustandsgröße des Brennstoffzellensystems.

Embodiments are explained in more detail below with reference to the accompanying drawings. Show it:

• 1 a schematic representation of a fuel cell system;
• 2 a schematic functional representation for determining the system state of the fuel cell system;
• 3 an example representation of a GRU entity as an RNN entity; and
• 4 a flowchart to illustrate the method for determining the system state variable of the fuel cell system.

Description of Embodiments

1 shows a schematic representation of a fuel cell system 1. The fuel cell system 1 has a fuel cell unit 2. FIG. The fuel cell unit 2 has a fuel cell 3 for providing electrical energy. The fuel cell 3 may correspond to a solid oxide fuel cell (SOFC: Solid Oxide Fuel Cell). Other cell technologies such as PLFC: Polymer Electrolyte Fuel Cell, MCFC: Molten Carbonate Fuel Cell and PAFC: Phosphoric Acid Fuel Cell can also be used accordingly.

The fuel cell 3 has an anode 5 and a cathode 6 . An electrode 7 is arranged between the anode 5 and the cathode 6, the anode 5 being separated from the cathode 6 by an electrolyte.

An oxidizing agent, in particular air or oxygen, can be supplied to the cathode 6 of the fuel cell unit 2 via a supply line 8 . The oxidizing agent is introduced via a valve 9 and compressed by means of an oxidizing compressor 10 . The oxidation compressor 10 can be variably adjusted in order to specify an oxygen mass flow. The oxidation compressor 10 can be designed as a constantly operated compressor with a variably adjustable throttle valve in order to variably adjust the oxygen mass flow as one of the manipulated variables.

To provide electrical energy, a fuel-containing synthesis gas, in particular a hydrogen-containing synthesis gas, is fed to the fuel cell unit 2 at the anode 5 via a fuel feed line 11 .

To generate the fuel-containing synthesis gas, the fuel cell unit 2 is preceded by a reformer device 12 . The reformer device 12 is connected to the fuel cell unit 2 or in particular to the anode 5 of the fuel cell 3 via the fuel supply line.

To generate the fuel gas-containing synthesis gas, the reformer device 12 is supplied with a mixture of natural gas or another reactant, in particular oxygen and/or steam, which is converted into the fuel gas-containing synthesis gas by reforming. The natural gas is compressed by a fuel compressor 13 and fed to the reformer device 12 via a fuel valve 14 . The fuel compressor 13 can be controlled variably in order to specify _{a fuel mass flow ṁ NG.} The fuel compressor 13 can be designed as a constantly operated compressor with a variably adjustable throttle valve in order to variably adjust the fuel mass flow as one of the manipulated variables.

A further reactant, water, in particular in the form of steam, is added to the natural gas downstream of the fuel compressor 13 through an exhaust gas recirculation system 15 .

The exhaust gas recirculation 15 of the fuel cell system 1 enables exhaust gas from the fuel cell 3, in the case shown an anode exhaust gas, to be at least partially returned to the input side of the reformer device 12 for mixing with the natural gas and the other reactants. For this purpose, the exhaust gas recirculation 15 can have a recirculation compressor 16 which is provided for the recirculation of water-containing anode exhaust gas to the reformer device 12 . The recirculation compressor 16 can be controlled variably in order to specify _{a recirculation mass flow ṁ Recy.} The recirculation compressor 16 can be designed as a constantly operated compressor with a variably adjustable throttle valve in order to variably adjust the recirculation mass flow as one of the manipulated variables.

During the reforming, long-chain alkanes are reformed completely and methane is partly reformed in the reformer device 12 and partly in the fuel cell 3 . Furthermore, unused synthesis gas can be fed back into the fuel cell by the reformer device 12 via the exhaust gas recirculation system 15 , which increases the fuel efficiency of the fuel cell system 1 . The various alkanes of the natural gas are fed into the reformer device 12 together with the water vapor and reformed there. Depending on the composition of the natural gas, a different amount of steam is required for optimized reforming.

The concentrations in the synthesis gas also depend on the composition of the natural gas used.

An internal state of the fuel cell system 1 is relevant for the operation of the fuel cell 3 . The state of the fuel cell system 1 is reflected in a system state variable, in particular in an oxygen-to-carbon ratio or in the fuel utilization rate of the fuel cell system 1 . Since the oxygen-to-carbon ratio or the fuel utilization rate cannot be measured directly or only with considerable effort, knowledge of the internal state of the fuel cell system 1 can help to operate it more effectively. For example, the system state variable can be determined as a function of the internal state of the fuel cell system 1 and used to control the operation of the fuel cell system 1 in a manner known per se.

In order to determine the system state variable, which is used as a parameter for the system state of the fuel cell system 1, measured variables must be recorded cyclically as operating state variables, which indicate the current operating state of the fuel cell system 1.

For example, using the temperature sensor, a volumetric mass flow sensor and a gas composition sensor, operating states of the fuel cell system can be detected as a temperature indication, one or more volume flow indications and a fuel gas composition, which make it possible to determine the system state variable based on an internal state. Furthermore, the operating state variables can include one or more manipulated variables provided for controlling the fuel cell system, in particular valve manipulated variables for controlling flow rates and the like.

Although the system state determined by the system state variable can be determined by a physical model as a function of the operating state variables, the measured operating state variables and/or the physical model are subject to errors, so that modeling an accumulation of errors that make it difficult to predict the system state of the fuel cell system 1 .

In 2 a functional circuit diagram is shown which specifies a system state model 20 for determining a system state variable x ⁽ⁱ⁾ , in particular in the form of a system state variable x ⁽ⁱ⁾ in the form of an oxygen-to-carbon ratio or a fuel utilization rate. The functional diagram shows a hybrid model with which the internal system state of the fuel cell system 1 can be determined.

The system status model 20 has a recursive character and has model blocks 21 which are executed at successive evaluation times i−1, i, i+1.

At each evaluation time i, the model block 21 receives as inputs a current operating ^{state variable vector y(i+1)} via the measured variables recorded at the current evaluation time and an internal state vector h ⁽ⁱ⁾ .

^{Alternatively, a number of internal state vectors h(in...i-1)} determined in the n previous evaluation times can also be taken into account.

The function of the model block 21 is executed for each evaluation time i based on the inputs in order ^{to determine a current system state variable x(i+1)} for a subsequent evaluation time i+1 and to update the internal state vector h ^(i+1). .

In detail, the system state model 20 includes a modeling block 22 that uses a physical model to determine a physically modeled system state variable x′. The physically modeled system state variable x' is determined using a physical model that depicts physical and/or chemical dependencies. This physical model determined based on the current operating variable vector y ^{(i + 1)} as the current temperature, the volumetric mass flow, gas composition, and the at least one manipulated variable and since the last evaluation time point to the current time of evaluation average volume of mass flow in the reformer apparatus 12 is a physically modeled system state variable x'.

The physically modeled system state variable x′ determined by the modeling block is supplied to a summing block 23 . The current system state variable x(i) corresponds to an absolute specification of the system state of the fuel cell system 1.

wobei K1, K2 zu parametrisierende Konstanten, ṁ_Recy einem Rückführungsmassenstrom und ṁ_NG einem Brennstoffmassenstrom angeben.In particular, in the case of x′ as the oxygen-carbon ratio φ, this can be determined according to the following physical model known from the prior art. $$\varphi =\frac{K2{\dot{m}}_{Recy}}{{\dot{m}}_{Recy}{\dot{m}}_{NG}+K1{{\displaystyle \dot{m}}}_{\mathrm{<figure-callout\; id="2"\; label="N\; G"\; filenames="DE102020208540A1\_0011.png"\; state="\{\{state\}\}">N</figure-callout>}G}^2}$$

where K1, K2 indicate constants to be parameterized, ṁ _Recy indicate a recirculation _{mass flow and ṁ NG} indicate a fuel mass flow.

wobei K3, K4 zu parametrisierenden Konstanten entsprechen.Alternatively, a fuel economy rate FU can be determined according to the following physical model known from the prior art: $$fu=\frac{K3K4{\dot{m}}_{NG}}{{\dot{m}}_{Recy}\left({\dot{m}}_{NG}+K4\right)+K3{{\displaystyle \dot{m}}}_{\mathrm{<figure-callout\; id="2"\; label="N\; G"\; filenames="DE102020208540A1\_0011.png"\; state="\{\{state\}\}">N</figure-callout>}G}^2}$$

where K3, K4 correspond to constants to be parameterized.

The physically modeled system state variable x', like the current operating ^{state variable vector y(i+1)} and the last determined internal state vector h ⁽ⁱ ), is supplied to an RNN unit 24 (RNN: recurrent neural network), in which a current internal -state vector h ^{(i+1) is} determined. The RNN unit 24 corresponds to a class of neural networks that are particularly well suited for processing temporal signal sequences and in particular for predicting temporal profiles. the RNN unit 24 can be configured, for example, as a GRU unit (GRU: Gated Recurrent Unit) or as an LSTM cell (Long-Short-Term-Memory) or as a variant thereof.

wird in einem optionalen Korrekturblock 25 verarbeitet, um eine Systemzustandsgrößenkorrektur K bereitzustellen und diese in dem Summierblock 23 der in dem vorangehenden Auswertungszeitpunkt i-1 ermittelten physikalisch modellierten Systemzustandsgröße x' zu addieren. Die Systemzustandsgrößenkorrektur K entspricht für obige physikalische Modell einer Korrekturgröße für das Sauerstoff-Kohlenstoff-Verhältnis ϕ oder einer Kraftstoffausnutzungsrate FU.The current internal state vector determined by the RNN unit 24 $\stackrel{\left(i\right)}{\mathrm{1..}T}$

is processed in an optional correction block 25 in order to provide a system state variable correction K and to add this in the summing block 23 to the physically modeled system state variable x′ determined in the preceding evaluation time i−1. For the above physical model, the system state variable correction K corresponds to a correction variable for the oxygen-carbon ratio φ or a fuel utilization rate FU.

The correction block 25 may be a fixed function that maps the internal state vector h ^{(i) to} the system state quantity correction K . The correction block 25 can be designed in particular as a neural network, preferably as a separate feed-forward neural network that ^{assigns the current internal state vector h(i+1) to} the system state variable correction K. In particular, in the case of an LSTM as the RNN unit 24, the correction block 25 can correspond to an “output gate” of the LSTM.

Alternatively, the RNN unit 24 can also be designed to ^{provide the system state variable correction K in addition to the current internal state vector h(i+1)} , so that the correction block 25 can be omitted.

After the function block 21 has been calculated, the current system ^{state variable x(i+1) is available at the output of the summing block 23 and the internal state vector h(i+1) is} available at the output of the RNN unit 24 for the evaluation time i+1. In the calculation at a next evaluation time i+2, these are ^{assumed as operating state variables together with the measured variables y(i+2) of} the next evaluation time i+2.

A GRU unit is trained as an example for the RNN unit 24 using known training methods, for example based on a backpropagation method. For this purpose, for a number k of predefined evaluation times Tk to T and a measured system state variable x ^(T) at an evaluation time T, an assumed or measured system state variable x ^(Tk) at an evaluation point in time Tk, the operating state variable vectors y ^{( Tk...T) are provided} as training variables and a corresponding recursive training method is carried out in a manner known per se in order to train the trainable parameters of the RNN.

$$r_t={\sigma }_g\left(W_r{y}_t+U_r{h}_{t-1}+b_r\right)$$

$$h_t=\left(1-z_t\right)°h_{t-1}+z_t°tanh\left(W_h{y}_t+U_h\left(r_t°h_{t-1}\right)+b_h\right)$$

$$z^{\left(i+1\right)}={\sigma }_g\left(W_z{y}^{\left(i+1\right)}+U_z{h}^{\left(i-1\right)}+b_z\right)$$

$$r^{\left(i+1\right)}={\sigma }_g\left(W_r{y}^{\left(i+1\right)}+U_r{h}^{\left(i\right)}+b_r\right)$$

$$h^{\left(i+1\right)}=\left(1-z^{\left(i+1\right)}\right)°h^{\left(i\right)}+z^{\left(i+1\right)}°tanh\left(W_n{y}^{\left(i+1\right)}+U_h\left(r_t°h^{\left(i\right)}\right)+b_h\right)$$

Wobei h_t dem Interner-Zustands-Vektor, y_t dem Betriebszustandsgrößenvektor, W, U, b Parametermatrizen bzw. Parametervektoren der GRU- Einheit und σ_g einer Sigmoid-Funktion entsprechen. 5 shows an exemplary representation of a GRU unit as an example of an RNN unit 24. The following function is carried out in the GRU unit: $${\mathrm{e.g}}_t={\sigma }_G\left(W_{\mathrm{e.g}}{y}_t+u_{\mathrm{e.g}}{H}_t+b_{\mathrm{e.g}}\right)$$

$${\mathrm{right}}_t={\sigma }_G\left(W_{\mathrm{right}}{y}_t+u_{\mathrm{right}}{H}_{t-1}+b_{\mathrm{right}}\right)$$

$$H_t=\left(1-{\mathrm{e.g}}_t\right)°H_{t-1}+{\mathrm{e.g}}_t°tanH\left(W_H{y}_t+u_H\left({\mathrm{right}}_t°H_{t-1}\right)+b_H\right)$$

$${\mathrm{e.g}}^{\left(i+1\right)}={\sigma }_G\left(W_{\mathrm{e.g}}{y}^{\left(i+1\right)}+u_{\mathrm{e.g}}{H}^{\left(i-1\right)}+b_{\mathrm{e.g}}\right)$$

$${\mathrm{right}}^{\left(i+1\right)}={\sigma }_G\left(W_{\mathrm{right}}{y}^{\left(i+1\right)}+u_{\mathrm{right}}{H}^{\left(i\right)}+b_{\mathrm{right}}\right)$$

$$H^{\left(i+1\right)}=\left(1-{\mathrm{e.g}}^{\left(i+1\right)}\right)°H^{\left(i\right)}+{\mathrm{e.g}}^{\left(i+1\right)}°tanH\left(W_n{y}^{\left(i+1\right)}+u_H\left({\mathrm{right}}_t°H^{\left(i\right)}\right)+b_H\right)$$

Where h _t corresponds to the internal state vector, y _{t to} the operating state variable vector, W, U, b to parameter matrices or parameter vectors of the GRU unit and σ _{g to} a sigmoid function.

^{The system state variable x(i+1)} determined in the evaluation for the current evaluation time i+1 is used to operate the fuel cell system. In particular, based on the system state variable, limit values for manipulated variables are defined, such as in M. Carre et al., “Feed-forward control of a solid oxide fuel cell system with anode offgas recycle”, Journal of Power Sources 282 (2015) , pages 498-510.

In 4 a flowchart to illustrate the method that is executed with the control device 10 is described.

In step S1, the operating state variables are measured at short sampling rates of, for example, <1 s, preferably less than 0.5 s, and if necessary made available together with the manipulated variables as an operating state variable vector.

In step S2, the current system state variable for the current evaluation time i+1 is determined from the operating state variables and the last determined internal state vector using the hybrid model.

In step S3, the operation of the fuel cell system is carried out ^{based on the determined current system state variable x(i).} A possible mode of operation is off, for example M.-Carre et al., "Feed-forward control of a solid oxide fuel cell system with anode offgas recycle", Journal of Power Sources 282 (2015), pages 498-510 described. Here, the values for the oxygen-carbon ratio φ or the fuel utilization rate FU determined via the model are used as limit values which must not be exceeded or fallen below when selecting an operating point of the fuel cell system.

QUOTES INCLUDED IN DESCRIPTION

This list of documents cited by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Non-patent Literature Cited

• M. Carre et al., "Feed-forward control of a solid oxide fuel cell system with anode offgas recycle", Journal of Power Sources 282 (2015) [0054]
• M.-Carre et al., "Feed-forward control of a solid oxide fuel cell system with anode offgas recycle", Journal of Power Sources 282 (2015), pages 498-510 [0058]

Claims (11)

Computer-implemented method for determining a system state variable (x ⁽ⁱ⁺¹⁾ ) in a fuel cell system (1) at successive evaluation times, the following steps being carried out at each current evaluation time (i+1): - determining (S2) one or more Operating state variables (y ⁽ⁱ⁺¹⁾ ) of the fuel cell system (1) at the current evaluation time (i+1); - Determining (S3) a current system ^{state variable (x (i + 1)} ) using a trained recurrent neural network depending on an internal state vector (h ⁽ⁱ⁾ ) of the recurrent neural network, an internal state of the fuel cell system (1) indicates at a previous evaluation time (i) and from the determined operating state variables (y ⁽ⁱ⁺¹⁾ ) at the current evaluation time (i+1); and - operation (S3) of the fuel cell system (1) depending on the currently determined system state variable (x ⁽ⁱ⁺¹⁾ ). procedure after claim 1 , wherein the determination (S3) of the current system state ^{variable (x (i+1)} ) is also carried out as a function of a progression of the operating state variables over a number of past evaluation times. procedure after claim 1 or 2 , wherein the operating state variables include at least one of the following variables: a generated output current, a temperature of a fuel cell unit (2) of the fuel cell system (1), one or more volume flows and a fuel composition at the inlet of the fuel cell system (1) and one or more manipulated variables, in particular valve manipulated variables for controlling flows, include or depend on them. Method according to one of the preceding claims, wherein the trained recurrent neural network is designed to, depending on the internal state vector (h ⁽ⁱ⁾ ) at the preceding evaluation time (i) and on the operating ^{state variables (y (i+1)} ) to determine a system state variable correction (K) at the current evaluation time (i+1) with which a ^{system state variable (x') modeled with the operating state variables (y (i+1)} ) using a physical model at the current evaluation time (i+1) is applied to determine the current system state variable (x ^{(i + 1)} ). procedure after claim 4 , the trained recurrent neural network taking into account the system state variable (x′) modeled using the physical model at the current evaluation time (i+1) in order to determine the system state variable correction (K). A method according to any one of the preceding claims, wherein the trained recurrent neural network comprises a GRU unit (24) or an LSTM unit. Procedure according to one of Claims 1 until 6 , wherein the fuel cell system (1) is operated as a function of the system state variable in that manipulated variables are limited to a value determined ^{by the current system state variable (x(i+1)).} Device, in particular control unit, for operating a fuel cell system (1), wherein a current system state variable (x ⁽ⁱ⁺¹⁾ ) is determined in the fuel cell system (1), wherein the device is designed to cyclically carry out the following steps: - determining (S2 ) of one or more operating state variables (y ⁽ⁱ⁺¹⁾ ) of the fuel cell system (1) at the current evaluation time (i+1); - Determining (S3) a current system ^{state variable (x (i + 1)} ) using a trained recurrent neural network depending on an internal state vector (h ⁽ⁱ⁾ ) of the recurrent neural network, an internal state of the fuel cell system (1) indicates at a previous evaluation time (i) and from the determined operating state variables (y ⁽ⁱ⁺¹⁾ ) at the current evaluation time (i+1); and - operation (S3) of the fuel cell system (1) depending on the currently determined system state variable (x ⁽ⁱ⁺¹⁾ ). Fuel cell system (1), comprising: - a fuel cell (3); - a reformer device (12); - An exhaust gas recirculation (15); - A sensor system for measuring measured variables as part of operating state variables; and - a device after claim 8 . Computer program which is set up to carry out all the steps of a method according to one of Claims 1 until 7 to execute. Electronic storage medium on which a computer program claim 10 is saved.

Images